An Introduction to Semiflows

CHAPMAN & HALL/CRC Monographs and Surveys in Pure and Applied Mathematics AN INTRODUCTION TO SEMIFLOWS 134 CHAPMAN ...

Author: Albert J. Milani | Norbert J. Koksch

23 downloads 732 Views 8MB Size Report

This content was uploaded by our users and we assume good faith they have the permission to share this book. If you own the copyright to this book and it is wrongfully on our website, we offer a simple DMCA procedure to remove your content from our site. Start by pressing the button below!

Report copyright / DMCA form

DOWNLOAD PDF

CHAPMAN & HALL/CRC Monographs and Surveys in Pure and Applied Mathematics

AN INTRODUCTION TO SEMIFLOWS

134

CHAPMAN & HALL/CRC Monographs and Surveys in Pure and Applied Mathematics Main Editors H. Brezis, Université de Paris R.G. Douglas, Texas A&M University A. Jeffrey, University of Newcastle upon Tyne (Founding Editor)

Editorial Board R. Aris, University of Minnesota G.I. Barenblatt, University of California at Berkeley H. Begehr, Freie Universität Berlin P. Bullen, University of British Columbia R.J. Elliott, University of Alberta R.P. Gilbert, University of Delaware R. Glowinski, University of Houston D. Jerison, Massachusetts Institute of Technology K. Kirchgässner, Universität Stuttgart B. Lawson, State University of New York B. Moodie, University of Alberta L.E. Payne, Cornell University D.B. Pearson, University of Hull G.F. Roach, University of Strathclyde I. Stakgold, University of Delaware W.A. Strauss, Brown University J. van der Hoek, University of Adelaide

CHAPMAN & HALL/CRC Monographs and Surveys in Pure and Applied Mathematics

AN INTRODUCTION TO SEMIFLOWS

Albert J. Milani Norbert J. Koksch

CHAPMAN & HALL/CRC A CRC Press Company Boca Raton London New York Washington, D.C.

134

C4584 disclaimer.fm Page 1 Wednesday, September 1, 2004 2:58 PM

Library of Congress Cataloging-in-Publication Data

This book contains information obtained from authentic and highly regarded sources. Reprinted material is quoted with permission, and sources are indicated. A wide variety of references are listed. Reasonable efforts have been made to publish reliable data and information, but the author and the publisher cannot assume responsibility for the validity of all materials or for the consequences of their use. Neither this book nor any part may be reproduced or transmitted in any form or by any means, electronic or mechanical, including photocopying, microfilming, and recording, or by any information storage or retrieval system, without prior permission in writing from the publisher. The consent of CRC Press does not extend to copying for general distribution, for promotion, for creating new works, or for resale. Specific permission must be obtained in writing from CRC Press for such copying. Direct all inquiries to CRC Press, 2000 N.W. Corporate Blvd., Boca Raton, Florida 33431. Trademark Notice: Product or corporate names may be trademarks or registered trademarks, and are used only for identification and explanation, without intent to infringe.

Visit the CRC Press Web site at www.crcpress.com © 2005 by Chapman & Hall/CRC Press No claim to original U.S. Government works International Standard Book Number 1-58488-458-4 Printed in the United States of America 1 2 3 4 5 6 7 8 9 0 Printed on acid-free paper

To Martina, Andrea, Daniel, Stephan and Claudia

Contents

1

2

Dynamical Processes 1.1 Introduction . . . . . . . . . . . . . . 1.2 Ordinary Differential Equations . . . 1.2.1 Well-Posedness . . . . . . . . 1.2.2 Regular and Chaotic Systems 1.2.3 Dependence on Parameters . . 1.2.4 Autonomous Equations . . . . 1.3 Attracting Sets . . . . . . . . . . . . 1.4 Iterated Sequences . . . . . . . . . . 1.4.1 Poincaré Maps . . . . . . . . 1.4.2 Bernoulli’s Sequences . . . . 1.4.3 Tent Maps . . . . . . . . . . . 1.4.4 Logistic Maps . . . . . . . . . 1.5 Lorenz’ Equations . . . . . . . . . . 1.5.1 The Differential System . . . 1.5.2 Equilibrium Points . . . . . . 1.6 Duffing’s Equation . . . . . . . . . . 1.6.1 The General Model . . . . . . 1.6.2 A Linearized Model . . . . . 1.7 Summary . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

1 1 4 6 8 10 11 16 20 22 24 27 30 32 33 33 36 36 38 40

Attractors of Semiflows 2.1 Distance and Semidistance . . . . . . . . . . 2.2 Discrete and Continuous Semiflows . . . . . 2.2.1 Types of Semiflows . . . . . . . . . . 2.2.2 Example: Lorenz’ Equations . . . . . 2.3 Invariant Sets . . . . . . . . . . . . . . . . . 2.3.1 Orbits . . . . . . . . . . . . . . . . . 2.3.2 Limit Sets . . . . . . . . . . . . . . . 2.3.3 Stability of Stationary Points . . . . . 2.3.4 Invariance of Orbits and ω-Limit Sets 2.4 Attractors . . . . . . . . . . . . . . . . . . . 2.4.1 Attracting Sets . . . . . . . . . . . . 2.4.2 Global Attractors . . . . . . . . . . . 2.4.3 Compactness . . . . . . . . . . . . . 2.5 Dissipativity . . . . . . . . . . . . . . . . . 2.6 Absorbing Sets and Attractors . . . . . . . .

. . . . . . . . . . . . . . .

. . . . . . . . . . . . . . .

. . . . . . . . . . . . . . .

. . . . . . . . . . . . . . .

. . . . . . . . . . . . . . .

. . . . . . . . . . . . . . .

. . . . . . . . . . . . . . .

. . . . . . . . . . . . . . .

. . . . . . . . . . . . . . .

. . . . . . . . . . . . . . .

. . . . . . . . . . . . . . .

. . . . . . . . . . . . . . .

41 41 42 42 46 49 50 52 55 59 64 64 65 67 68 70

. . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . .

vii

viii . . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

70 72 73 74 79 81 83 84 86

Attractors for Semilinear Evolution Equations 3.1 PDEEs as Dynamical Systems . . . . . . . . . . . . 3.1.1 The Model IBV Problems . . . . . . . . . . 3.1.2 Construction of the Attractors . . . . . . . . 3.2 Functional Framework . . . . . . . . . . . . . . . . 3.2.1 Function Spaces . . . . . . . . . . . . . . . 3.2.2 Orthogonal Bases . . . . . . . . . . . . . . . 3.2.3 Finite Dimensional Subspaces . . . . . . . . 3.3 The Parabolic Problem . . . . . . . . . . . . . . . . 3.3.1 Step 1: The Solution Operator . . . . . . . . 3.3.2 Step 2: Absorbing Sets . . . . . . . . . . . . 3.3.3 Step 3: Compactness of the Solution Operator 3.3.4 Step 4: Conclusion . . . . . . . . . . . . . . 3.3.5 Backward Uniqueness . . . . . . . . . . . . 3.4 The Hyperbolic Problem . . . . . . . . . . . . . . . 3.4.1 Step 1: The Solution Operator . . . . . . . . 3.4.2 Step 2: Absorbing Sets . . . . . . . . . . . . 3.4.3 Step 3: Compactness of the Solution Operator 3.4.4 Step 4: Conclusion . . . . . . . . . . . . . . 3.4.5 Attractors via α-Contractions . . . . . . . . 3.5 Regularity . . . . . . . . . . . . . . . . . . . . . . 3.6 Upper Semicontinuity of the Global Attractors . . .

. . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . .

89 89 89 92 95 95 97 98 99 100 105 106 107 108 111 112 114 116 121 121 123 132

Exponential Attractors 4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . 4.2 The Discrete Squeezing Property . . . . . . . . . . . . 4.2.1 Orthogonal Projections . . . . . . . . . . . . . . 4.2.2 Squeezing Properties . . . . . . . . . . . . . . . 4.2.3 Squeezing Properties and Exponential Attractors 4.2.4 Proof of Theorem 4.5 . . . . . . . . . . . . . . . 4.3 The Parabolic Problem . . . . . . . . . . . . . . . . . . 4.3.1 Step 1: Absorbing Sets in X1 . . . . . . . . . . . 4.3.2 Step 2: The Discrete Squeezing Property . . . . 4.4 The Hyperbolic Problem . . . . . . . . . . . . . . . . . 4.4.1 Step 1: Absorbing Sets in X1 . . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

135 135 137 137 138 139 141 143 143 144 147 147

2.7

2.8 2.9

3

4

2.6.1 Attractors of Compact Semiflows . 2.6.2 A Generalization . . . . . . . . . . Attractors via α-Contractions . . . . . . . 2.7.1 Measuring Noncompactness . . . . 2.7.2 A Route to α-Contractions . . . . . Fractal Dimension . . . . . . . . . . . . . A Priori Estimates . . . . . . . . . . . . . 2.9.1 Integral and Differential Inequalities 2.9.2 Exponential Inequality . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

ix . . . . . . . . .

. . . . . . . . .

149 152 152 153 155 159 161 162 175

Inertial Manifolds 5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.2 Definitions and Comparisons . . . . . . . . . . . . . . . . . . . 5.2.1 Lipschitz Manifolds and Inertial Manifolds . . . . . . . . 5.2.2 Inertial Manifolds and Exponential Attractors . . . . . . . 5.2.3 Methods of Construction of the Inertial Manifold . . . . . 5.3 Geometric Assumptions on the Semiflow . . . . . . . . . . . . . 5.3.1 The Cone Invariance Property . . . . . . . . . . . . . . . 5.3.2 Decay and Squeezing Properties . . . . . . . . . . . . . . 5.3.3 Consequences of the Decay Property . . . . . . . . . . . 5.4 Strong Squeezing Property and Inertial Manifolds . . . . . . . . 5.4.1 Surjectivity and Uniform Boundedness . . . . . . . . . . 5.4.2 Construction of the Inertial Manifold . . . . . . . . . . . 5.5 A Modification . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.5.1 The Modified Strong Squeezing Property . . . . . . . . . 5.5.2 Consequences of the Modified Strong Squeezing Property 5.5.3 Construction of the Inertial Manifold, 2 . . . . . . . . . . 5.5.4 Comparison of the Squeezing Properties . . . . . . . . . . 5.6 Inertial Manifolds for Evolution Equations . . . . . . . . . . . . 5.6.1 The Evolution Problem . . . . . . . . . . . . . . . . . . . 5.6.2 The Spectral Gap Condition . . . . . . . . . . . . . . . . 5.6.3 The Strong Squeezing Properties . . . . . . . . . . . . . . 5.6.4 Uniform Boundedness and Surjectivity . . . . . . . . . . 5.7 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.7.1 Semilinear Heat Equations . . . . . . . . . . . . . . . . . 5.7.2 Semilinear Wave Equations . . . . . . . . . . . . . . . . 5.8 Semilinear Evolution Equations in One Space Dimension . . . . 5.8.1 The Parabolic Problem . . . . . . . . . . . . . . . . . . . 5.8.2 Absorbing Sets . . . . . . . . . . . . . . . . . . . . . . . 5.8.3 Adjusting the Nonlinearity . . . . . . . . . . . . . . . . . 5.8.4 The Inertial Manifold . . . . . . . . . . . . . . . . . . . . 5.8.5 The Hyperbolic Perturbation . . . . . . . . . . . . . . . . 5.8.6 Concluding Remarks . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

177 177 179 179 183 185 189 189 191 193 195 195 197 201 201 203 204 206 208 208 209 212 215 218 219 220 229 229 230 234 235 238 239

4.5

4.6 5

4.4.2 Step 2: The Discrete Squeezing Property Proof of Theorem 4.4 . . . . . . . . . . . . . . . 4.5.1 Outline . . . . . . . . . . . . . . . . . . 4.5.2 The Cone Property . . . . . . . . . . . . 4.5.3 The Basic Covering Step . . . . . . . . . 4.5.4 The First and Second Iterates . . . . . . . 4.5.5 The General Iterate . . . . . . . . . . . . 4.5.6 Conclusion . . . . . . . . . . . . . . . . Concluding Remarks . . . . . . . . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

x 6

7

Examples 6.1 Cahn-Hilliard Equations . . . . . . . . . . . . . . . . 6.1.1 Introduction . . . . . . . . . . . . . . . . . . . 6.1.2 The Cahn-Hilliard Semiflows . . . . . . . . . 6.1.3 Absorbing Sets . . . . . . . . . . . . . . . . . 6.1.4 The Global Attractor . . . . . . . . . . . . . . 6.1.5 The Exponential Attractor . . . . . . . . . . . 6.1.6 The Inertial Manifold . . . . . . . . . . . . . . 6.2 Beam and von Kármán Equation . . . . . . . . . . . 6.2.1 Functional Framework and Notations . . . . . 6.2.2 The Beam Equation Semiflow . . . . . . . . . 6.2.3 Absorbing Sets . . . . . . . . . . . . . . . . . 6.2.4 The Global Attractor . . . . . . . . . . . . . . 6.2.5 The Exponential Attractor . . . . . . . . . . . 6.2.6 Inertial Manifold . . . . . . . . . . . . . . . . 6.2.7 von Kármán Equations . . . . . . . . . . . . . 6.3 Navier-Stokes Equations . . . . . . . . . . . . . . . . 6.3.1 The Equations and their Functional Framework 6.3.2 The 2-Dimensional Navier-Stokes Semiflow . . 6.3.3 Absorbing Sets and Attractor . . . . . . . . . . 6.3.4 The Exponential Attractor . . . . . . . . . . . 6.4 Maxwell’s Equations . . . . . . . . . . . . . . . . . 6.4.1 The Equations and their Functional Framework 6.4.2 The Quasi-Stationary Maxwell Semiflow . . . 6.4.3 Absorbing Sets and Attractors . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . .

241 241 242 244 247 250 251 254 261 261 262 263 267 269 271 272 272 272 275 276 278 280 281 285 288

A Nonexistence Result for Inertial Manifolds 7.1 The Initial-Boundary Value Problem . . . . . 7.2 Overview of the Argument . . . . . . . . . . 7.3 The Linearized Problem . . . . . . . . . . . 7.4 Inertial Manifolds for the Linearized Problem 7.5 C1 Linearization Equivalence . . . . . . . . 7.6 Perturbations of the Nonlinear Flow . . . . . 7.7 Asymptotic Properties of the Perturbed Flow 7.8 The Nonexistence Result . . . . . . . . . . . 7.9 Proof of Proposition 7.17 . . . . . . . . . . . 7.10 The C1 Linearization Equivalence Theorems 7.10.1 Equivalence for a Single Operator . . 7.10.2 Equivalence for Groups of Operators

Appendix: Selected Results from Analysis A.1 Ordinary Differential Equations . . . . . A.1.1 Classical Solutions . . . . . . . . A.1.2 Generalized Solutions . . . . . . A.1.3 Stability for Autonomous Systems

. . . .

. . . .

. . . . . . . . . . . .

. . . . . . . . . . . .

. . . . . . . . . . . .

. . . . . . . . . . . .

. . . . . . . . . . . .

. . . . . . . . . . . .

. . . . . . . . . . . .

. . . . . . . . . . . .

. . . . . . . . . . . .

. . . . . . . . . . . .

. . . . . . . . . . . .

. . . . . . . . . . . .

291 291 293 295 298 303 304 307 309 310 316 316 320

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

323 323 323 324 325

xi A.2 Linear Spaces and their Duals . . . . . . . . . . . . . . . A.2.1 Orthonormal Bases in Hilbert spaces . . . . . . . . A.2.2 Dual Spaces and the Hahn-Banach Theorem . . . . A.2.3 Linear Operators in Banach Spaces . . . . . . . . A.2.4 Adjoint of a Bounded Operator . . . . . . . . . . . A.2.5 Adjoint of an Unbounded Operator . . . . . . . . A.2.6 Gelfand Triples of Hilbert Spaces . . . . . . . . . A.2.7 Linear Operators in Gelfand Triples . . . . . . . . A.2.8 Eigenvalues of Compact Operators . . . . . . . . . A.2.9 Fractional Powers of Positive Operators. . . . . . A.2.10 Interpolation Spaces . . . . . . . . . . . . . . . . A.2.11 Differential Calculus in Banach Spaces . . . . . . A.3 Semigroups of Linear Operators . . . . . . . . . . . . . . A.3.1 General Results . . . . . . . . . . . . . . . . . . . A.3.2 Applications to PDEs . . . . . . . . . . . . . . . . A.4 Lebesgue Spaces . . . . . . . . . . . . . . . . . . . . . . A.4.1 The Spaces L p (Ω ) . . . . . . . . . . . . . . . . . A.4.2 Inequalities . . . . . . . . . . . . . . . . . . . . . A.4.3 Other Properties of the Spaces L p (Ω ) . . . . . . . A.5 Sobolev Spaces of Scalar Valued Functions . . . . . . . . A.5.1 Distributions in Ω . . . . . . . . . . . . . . . . . A.5.2 The Spaces Hm (Ω ), m ∈ N . . . . . . . . . . . . . A.5.3 The Spaces Hs (Ω ), s ∈ R≥0 . . . . . . . . . . . . A.5.4 The Spaces Hs0 (Ω ), s ∈ R≥0 , and Hs (Ω ), s ∈ R<0 A.5.5 The Laplace Operator . . . . . . . . . . . . . . . A.6 Sobolev Spaces of Vector Valued Functions . . . . . . . . A.6.1 Lebesgue and Sobolev Spaces . . . . . . . . . . . A.6.2 The Intermediate Derivatives Theorem . . . . . . . A.7 The Spaces H(div, Ω ) and H(curl, Ω ) . . . . . . . . . . . A.7.1 Notations . . . . . . . . . . . . . . . . . . . . . . A.7.2 The Space H(div, Ω ) . . . . . . . . . . . . . . . . A.7.3 The Space H(curl, Ω ) . . . . . . . . . . . . . . . A.7.4 Relations between H(div, Ω ) and H(curl, Ω ) . . . A.8 Almost Periodic Functions . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

328 328 329 330 333 335 335 336 338 340 342 344 345 345 347 349 349 350 351 352 353 353 355 357 358 361 361 362 363 364 364 365 367 371

Bibliography

373

Index

381

Nomenclature

385

Preface

1. In these notes we present some introductory material on a particular class of dynamical systems, called SEMIFLOWS. This class includes, but is not restricted to, systems that are defined, or modelled, by certain types of differential equations of evolution (DEEs in short). Our purpose is to introduce, in a relatively self-contained manner, the basic results of the theory of dynamical systems that can be naturally extended and applied to study the asymptotic behavior of the solutions of the DEEs we consider. Equations of evolution include ordinary differential equations (ODEs in short), partial differential equations of evolution (PDEEs in short), and other types of equations, such as, for instance, stochastic or difference equations. As such, they provide natural examples of dynamical systems, since one of the independent variables (usually called “time”) plays a different role than the other variables (which in some situations may be called “space” variables). Thus, in this context, the heat and wave equations are considered as prototypical examples of PDEEs, while elliptic equations such as Laplace’s equation are not considered as evolution equations, because in these equations all the variables have the same role. Here, we make the further distinction that “time” evolves continuously; thus, we do not consider stochastic equations, nor, except for some introductory examples, discrete systems (where “time” varies along a sequence). 2. One of the major goals of the theory of dynamical systems is the study of the evolution of a system, with the purpose of predicting, as accurately as possible, the behavior of the system as time becomes large. A quite general feature of the systems we consider, which is shared with other systems, is a property called DISSIPATIVITY . Loosely speaking, this property can be described by the fact that all solutions of these systems eventually enter, and remain, in a bounded set, called ABSORB ING SET. Thus, the evolution of the solutions of the system can be studied in this set; as a result, the long time behavior of the system can be described by means of certain subsets of the absorbing set. Among these, we shall consider three types of sets, called respectively ATTRACTORS, EXPONENTIAL ATTRACTORS, and INERTIAL MANIFOLDS . (Exponential attractors are sometimes also known as INERTIAL SETS .) We will present the fundamental properties of these sets, and then proceed to show the existence of some of these sets for a number of dynamical systems, generated by fairly well known physical models. In particular, we shall consider in full detail two particular PDEEs of evolution: a semilinear version of the heat equation, and a corresponding version of the dissipative wave equation. These examples allow us to illustrate the most important features of the theory of semiflows, and to provide a sort of “template” that can then be applied, in a more or less straightforward fashion,

xiii

xiv to the analysis of other models, with the help of the many specialized references that exist in the literature. 3. Even a quick survey of much of the existing literature on dynamical systems, both at the introductory and the specialized level, reveals that the notion of “dynamical system” is used with many different meanings, according to the specific point of view of the authors. At the opposite extreme, this notion may well be not defined at all. In these notes, we do not attempt to give a general definition of dynamical system; rather, we confine ourselves to a special class of systems, properly known as CONTINUOUS , SEMI - DYNAMICAL SYSTEMS , or CONTINUOUS SEMIFLOWS. Here, the term “continuous” is used to distinguish these systems from DISCRETE ones, where only a sequence of successive time values are considered, and “semi-” refers to the fact that time evolves, i.e. we only consider nonnegative values of the time variable. For brevity, we shall refer to these systems as SEMIFLOWS (their precise definition is given in section 2.2). In the introductory chapter 1, we consider more general TWO - PARAMETER SEMIFLOWS or DYNAMICAL PROCESSES, which allows us to include some nonautonomous difference or differential equations as generators of dynamical systems. However, when our presentation can proceed in a more discursive way, and rigor is not an issue, we conform to the common use and adopt the general term “dynamical system”. 4.

In general, we say that an ODE defines a semiflow if the corresponding C AUCHY is globally well posed, in the sense we define in section 1.2.1. We can extend this definition to semiflows defined by PDEEs, by interpreting the PDEE as an abstract ODE in a suitable Banach space X (see remark 3.2 in chapter 3). This is a generalization of the usual interpretation of a system of ODEs as a single differential equation in the Banach space X = Rn , or in more general finite dimensional vector spaces, and explains the qualification of the systems generated by PDEEs as “infinite dimensional” ones, since in this case X is in general no longer a finite dimensional space. Examples of PDEEs that can be put in such abstract form are: the Navier-Stokes equations, the Kuramoto-Sivashinski equations, the “original” Burger’s equation, the Chafee-Infante and Cahn-Hilliard reaction-diffusion equations, the Korteweg-de Vries and the Maxwell equations (see chapter 6). Indeed, many basic notions and results in the theory of the asymptotic behavior of infinite dimensional dissipative dynamical systems trace their origin in the study of the NavierStokes equations of fluid dynamics, and have been inspired by a detailed analysis of both the qualitative properties of their solutions, and their behavior with respect to numerical computations. PROBLEM

5. Not surprisingly, much of the general terminology in the theory of dynamical systems, as well as the general spirit of its qualitative results, borrows directly from the qualitative theory of ODEs in Rn . For example, we shall need to recall some basic results on stability, equilibrium points, periodic orbits, ω-limit sets, etc. On the other hand, in an effort to keep these notes within a reasonable length, we shall

xv be forced to not discuss many other important topics. In particular, we regretfully do not include any result on bifurcation theory. Among the many excellent and fairly complete references on the qualitative theory of ODEs, including ODEs as dynamical systems, we refer for example to Hirsch and Smale, [HS93], Jordan and Smith, [JS87], Perko, [Per91], Amann, [Ama90], and Verhulst, [Ver90]. A few other references, specifically on dynamical systems, are listed in the bibliography. Since so many articles and books are continually being published, it is almost impossible to compile an exhaustive list of references; on the other hand, an internet search can provide all necessary updated references on any particular topic. 6. These notes have their origin in a series of graduate seminars we held at the Universities of Dresden, Wisconsin-Milwaukee and Tsukuba. Most of the material we cover is relatively well known, although some of the results we present, in particular on the existence of an exponential attractor and of an inertial manifold for semilinear dissipative wave equations, even if not entirely new, do not seem to enjoy the recognition we feel they deserve. In part, our intention in writing these notes is to be of some help to “beginners” in the area of infinite dimensional dynamical systems; that is, anyone who, having a solid background in the classical theory of ODEs and some knowledge of functional analysis in Sobolev spaces, wishes to proceed to the study of examples of semiflows arising from DEEs, but may need some “smoothing into” the subject, before turning to more general introductory texts, such as those of Temam, [Tem88], the cycle of lectures by Oleinik, [Ole96], or, most recently, SellYou, [SY02], and Robinson, [Rob01]. We also hope that these notes may serve as a ready reference to researchers in more applied fields, who feel the need for a clear presentation of the background material and results that are necessary for the study of the specific systems they are interested in. To this end, we have tried to “build up” the material in as careful and gradual progression as possible, with the goal of presenting the main topics (in particular, the construction of the exponential attractor and the inertial manifold), with a larger degree of detail than generally found in most sources in the literature. If successful, our effort should put the reader in a better position to refer to more specific texts on global attractors, exponential attractors, and inertial manifolds, such as, respectively, the books by Babin and Vishik, [BV92], Eden, Foias, Nicolaenko and Temam, [EFNT94], and Constantin, Foias, Nicolaenko and Temam, [CFNT89]. 7. These notes are organized as follows. As an introduction to the main ideas of the abstract theory of semiflows, in chapter 1 we present some well known and well studied examples of finite dimensional dynamical systems, generated by such celebrated ODEs as Duffing’s equations and Lorenz’ equations. In chapter 2 we introduce the general definitions of SEMIFLOWS and their GLOBAL ATTRACTORS, and we present two sufficient conditions that guarantee the existence of the attractor under different assumption on the asymptotic properties of the semiflow. We also describe an alternate construction of the attractor, based on the idea of α-contracting maps. In chapter 3 we apply these results to show that the semiflows generated by

xvi two types of semilinear dissipative evolution PDEEs (one parabolic and the other hyperbolic) admit a global attractor in a suitable space of weak solutions. In chapter 4 we briefly develop the theory of EXPONENTIAL ATTRACTORS, and apply this theory to the models of PDEEs considered in chapter 3. In chapter 5 we present Hadamard’s GRAPH TRANSFORMATION METHOD for the construction of an INER TIAL MANIFOLD , and apply this method to a one-dimensional version of the PDEEs considered in chapter 3. In chapter 6, we consider a number of other dynamical systems, generated by PDEEs that model various mathematical physics problems, and briefly show how the methods developed in the previous chapters can be applied. In chapter 7 we present a result, due to Mora and Solà-Morales, on the nonexistence of inertial manifolds for the semiflow generated by a one-dimensional version of the hyperbolic model of PDEE considered in chapter 3. Finally, in the Appendix we collect, for the reader’s convenience, a list of various definitions and results from the classical theory of ODEs and PDEs, functional and nonlinear analysis, semigroup theory and Lebesgue-Sobolev spaces, that we use in these notes, and provide at least one reference for each of the definitions and theorems we state. Acknowledgements. Both authors have been partially supported by the Alexander von Humboldt Stiftung. The second author also had partial support from the Japanese Society for the Promotion of Sciences and the Institute of Mathematics of Fudan University in Shanghai, and is grateful to his colleagues at the Institut für Analysis of the TU Dresden for their warm hospitality. We are also greatly indebted to Professor Songmu Zheng of Fudan University for very kind and stimulating discussion, and to the anonymous referees of this book for various helpful suggestions and corrections.

Chapter 1 Dynamical Processes

In this chapter we introduce the definition of DYNAMICAL PROCESS, and the main ideas of the theory of dynamical systems that we want to investigate. We illustrate these ideas by examining some simple examples of dynamical processes generated by finite systems of ODEs and by iterated maps.

1.1 Introduction 1. Roughly speaking, the theory of dynamical systems studies mathematical models of physical “systems” which evolve in time from a “state” which is known at an initial moment; more specifically, how the evolution of a system depends, or is influenced by, its initial state. The changing density of a population from a known number of individuals (e.g., sharks in a regional sea; bacteria in an infected organism; prey-predator models); the changing of weather patterns in a particular region; the spreading of a rumor; the vapor trail in the wake of an airplane; the propagation of a fire: all these would be examples of dynamical systems. To study the evolution of a system, we assume that its state at each time t can be described generally by means of a function t 7→ u(t), where the independent “time” variable t is measured in a parameter set T ⊂ R, and the corresponding dependent variable is in a set X , called STATE SPACE. We also assume that the state u(t) of the system at any given time t depends not only on the value of t, but also on its initial configuration, i.e. on the value u0 of the system at a previous time t0 , with u0 and t0 given or known. A natural goal of the theory is then to study the dependence of the state u ∈ X on the time t ∈ T and the INITIAL VALUES u0 ∈ X , t0 ∈ T . In particular, we can think of a dynamical system as a way of transforming an initial state u0 into a family of subsequent states u(t), parametrized by t ∈ T . We shall indeed assume that there is a specified functional dependence of u ∈ X from u0 ∈ X and t, t0 ∈ T , described by a map (t,t0 , u0 ) 7→ u(t,t0 , u0 ) .

(1.1)

By specifying certain properties of this map, we come to a definition of a special kind of dynamical systems.

1

2

1

Dynamical Processes

DEFINITION 1.1 Let X be an arbitrary set, and T be one of the sets N, Z, R≥0 or R, where R≥0 := [0, +∞[. Set T∗2 := {(t, τ) ∈ T × T : t ≥ τ} . A TWO - PARAMETER SEMIFLOW, or DYNAMICAL PROCESS in X is a family S = (S(t, τ))(t,τ)∈T∗2 of maps S(t, τ) : X → X , which satisfies the following conditions: ∀t ∈ T :

S(t,t) = IX

(1.2)

(the identity in X ), and ∀ t1 ,t2 ,t3 ∈ T :

S(t1 ,t2 )S(t2 ,t3 ) = S(t1 ,t3 ) .

(1.3)

The following are familiar examples of dynamical processes. Example 1.2 Let X = R and T = R. Let f be a continuous function on R, and S = (S(t, τ))(t,τ)∈T∗2 be the family of maps S(t, τ) : R → R defined by Z t S(t, τ)x := exp f (s) ds x, x ∈ R. (1.4) τ

Then, S is a dynamical process in R. Indeed, verification of (1.2) and (1.3) is immediate. Example 1.3 Let X = Rn , and A be an n × n matrix. Then, the family T = (etA )t ∈R of the exponentials of the matrices tA is a linear semigroup in X (see section A.3). Consequently, the family S defined by S(t, τ) := e(t −τ)A ,

(t, τ) ∈ R2 ,

is a dynamical process. Note that, in these examples, each map S(t, τ) is linear; as we shall see, this needs not be the case in general. According to definition 1.1, a dynamical process S on a set X consists of a family of transformations of X into itself, each defined by the map (1.1), that is, X 3 u0 7→ u(t, τ, u0 ) =: S(t, τ)u0 ∈ X .

(1.5)

We are then mainly interested in the dependence of the map t 7→ S(t,t0 )u0 on the “initial values” t0 and u0 or, sometimes, on u0 only, for fixed t0 . Of course, this requires X to have some kind of topological structure, and we shall remove the provisional nature of definition 1.1, supplementing it by a number of continuity conditions on

1.1

Introduction

3

the maps S(t, τ) on X , and of the map (t, τ) 7→ S(t, τ). In particular, as the examples we cited above indicate, we are often interested in being able to describe, or determine, the evolution of a given system “in the future”. This question can be clearly related to the asymptotic properties, as t → +∞ (in T ), of the map defined in (1.1). Because of (1.5), we are then naturally led to relate the asymptotic behavior of the function u to some suitable properties of the corresponding dynamical process S, defined by (1.5). For example, a possible question would be to determine all the values (u0 ,t0 ) ∈ X × T such that the limit lim S(t,t0 )u0 =: L(u0 )

t →+∞

(1.6)

exists, for a fixed t0 . As an illustration, if S is the dynamical process defined in (1.4), the limit in (1.6) exists for all u0 ∈ R if f is bounded above by a negative constant. Note that, since in this case L(u0 ) = 0 for all u0 ∈ R, this limit is actually independent of the initial value u0 . Another, related question would be to study the properties of the map u0 7→ L(u0 ) defined by (1.6). 2. In these notes, we assume that the state space X is at least a Banach space (on R), and the underlying time parameter set T will be either N or Z, in which case we call the system DISCRETE, or R≥0 or R, in which case we call the system CONTIN UOUS . In this chapter we propose to give a first idea of the nature of the questions, related to the long time behavior of dynamical processes, that we want to investigate. To do so, we consider some introductory examples of discrete dynamical processes, generated by iterated maps, and of continuous dynamical processes, generated by finite systems of ODEs. In these cases, the Banach space X has finite dimension, and the corresponding dynamical process is also called FINITE DIMENSIONAL. In chapter 3 we will instead consider INFINITE DIMENSIONAL dynamical processes, generated by PDEs of evolution. In this case, the space X is infinite dimensional; specifically, a space of functions of some “space” variables, defined on a domain of Rn . One can find a large amount of examples of this type of systems in specialized texts, such as Jordan-Smith, [JS87], Marsden-McCracken, [MM76], GuckenheimerHolmes, [GH83], Moon, [Moo92], Alligood-Sauer-Yorke, [ASY96], and many others. Among the most studied examples, we recall the models known as Duffing’s equation, the logistic equation, the Lorenz system, and Hénon’s horseshoe map. Most of these also illustrate another major goal of the theory of the dynamical systems, which, regretfully, we cannot pursue because of the introductory character of these notes. Namely, all these systems depend on various numerical parameters, and the influence of these parameters on the long time behavior of the system exhibits some striking phenomena, and unexpected similarities among these systems. In particular, even if the parameters are allowed to vary in a continuous fashion, and even if for a certain range of the parameters the evolution of the system seems to be quite “regular”, for other parameter ranges a number of other, totally new qualitative phenomena unexpectedly appear. Examples of such phenomena are BIFURCATIONS (see e.g. Marsden-McCracken, [MM76]), F EIGENBAUM CASCADES (e.g. for the logistic

4

1

Dynamical Processes

map described in section 1.4.4; see e.g. Moon, [Moo92], or Feigenbaum, [Fei78]), and HORSESHOE MAPS (e.g. for Hénon’s map, whose iterations converge to a set of so-called FRACTAL type; see Hénon-Pomeau, [HP76]).

1.2 Ordinary Differential Equations 1. As a first example of dynamical processes, we consider continuous systems generated by an evolution equation of the form u˙ = F(t, u) ,

(1.7)

where F : R×X → X is a continuous function on a Banach space X . In this case, we take T = R or T = R≥0 . If X is finite dimensional, (1.7) is equivalent to a system of ODEs in Rn , where n is the dimension of X . An example is the system of m coupled pendulums on the same vertical plane: In this case, if θ1 , . . . , θm denote the angles of each pendulum with respect to the vertical, then u = (θ1 , θ˙1 , . . . , θm , θ˙m ) and X = R2m . We shall, however, be more interested in the case when the dimension of X is infinite, and (1.7) represents a PDEE, interpreted as an abstract evolution equation in X . An example is the semilinear heat equation ut = ∆u + f (u)

(1.8)

in a domain Ω ⊂ Rn , with appropriate boundary conditions. In this case, the space X is a space of functions defined on Ω ; for example, we can consider the Lebesgue space L2 (Ω ), or the Sobolev space H10 (Ω ), or the Hölder space C0,α (Ω ). We can then interpret PDEEs like (1.8) as abstract ODEs in X by means of the following ˜ ∈X natural identification. If u is a solution of (1.8), we define a function t 7→ u(t) by (u(t)) ˜ (x) := u(t, x) , x∈Ω; that is, we consider for each t the image u(t) ˜ ∈ X as a function of the space variable x. It is common practice to identify u and u, ˜ introducing the notation u(t, ·) := u(t) ˜ , which we shall often adopt. 2. We assume that, in accord with the classical (Newtonian) theory, equation (1.7) is deterministic, in the sense that the knowledge of the initial values (t0 , u0 ) (and, of course, of F) uniquely determines a solution u, defined for all “future times”, of the Cauchy problem corresponding to (1.7), that is ( u˙ = F(t, u) , (1.9) u(t0 ) = u0 .

1.2

Ordinary Differential Equations

5

More precisely, we assume that under sufficient assumptions on the function F, there is a unique function u ∈ C([t0 , +∞[; X ), which satisfies the Cauchy problem (1.9), either in the classical sense (if e.g. u is also in C1 ([t0 , +∞[; X )), or in a generalized sense (e.g. almost everywhere in t, or in distributional sense). This solution is typically determined at first only locally in time, that is, on a neighborhood ]t0 −α,t0 +β [ of t0 , and then extended uniquely to a function, which is defined at least on the unbounded interval ]t0 − α, +∞[, and solves problem (1.9) on the whole interval [t0 , +∞[. We usually denote this extended function again by u. Of course, in some cases the local solution u could also be extended to the left of t0 − α; however, since in the context of evolution problems we are mostly interested in what happens in “the future”, we will generally not be too concerned about the possibility of extending u to the left of t0 . (We also note in passing that, when trying to do so, we sometimes meet additional problems, such as the lack of backward uniqueness.) Thus, when in the sequel we use the term “global solution”, we always refer to solutions that are defined globally at least to the right of t0 , i.e. for all t ≥ t0 . Clearly, the possibility of extending a local solution to a global one must in general be proven for each specific problem. This can be done in different ways; a common one is to show that any local solution satisfies a number of so-called A PRIORI ES TIMATES . These are bounds on the solution which are independent of the particular time interval where the solution is defined, and therefore allow us to extend any local solution uniquely to a global one, by means of a repeated application of the local existence result. 3. Having thus established a unique solution u of the Cauchy problem (1.9) for all choices of initial values (t0 , u0 ), we are then interested in the asymptotic behavior of u(t) as t → +∞. More specifically, we would like to understand how this behavior is determined (if at all) by the initial values u0 and t0 (or, in some cases, by u0 only). To this end, it is convenient to introduce more proper notations. To emphasize that the solution u depends not only on t, but also on the initial values (t0 , u0 ), we consider u as a function defined on R ×R × X , with values in X , and write u(t,t0 , u0 ) to indicate the image of the point (t,t0 , u0 ) by u. Next, we realize that the solution of the Cauchy problem (1.9) defines a family S := (S(t,t0 ))(t,t0 )∈Θ , Θ := {(t, s) : t ≥ s} ,

(1.10)

of operators S(t,t0 ) : X → X , parametrized by the pair (t,t0 ) in the half-plane Θ . Each operator S(t,t0 ) is defined by S(t,t0 )u0 := u(t,t0 , u0 ) ,

u0 ∈ X .

(1.11)

This family S is called the family of SOLUTION OPERATORS associated to (or, defined by) equation (1.7). Standard uniqueness theorems on solutions of the Cauchy problem (1.9) can then be used to verify that S satisfies conditions (1.2) and (1.3) of definition 1.1; hence, S is a dynamical process on X . We say that S is GENERATED by problem (1.9). We also say that the map t 7→ S(t,t0 )u0 defined in (1.11) is a MO TION of the dynamical process S, corresponding to the initial values (t0 , u0 ), and the

6

1

Dynamical Processes

image of this motion is an ORBIT of the system (a more precise definition of motions and orbits will be given in section 1.2.4). Example 1.4 The Cauchy problem (

y˙ = f (t)y , y(t0 ) = y0

(1.12)

generates the dynamical process S defined in (1.4). Indeed, (1.12) has the unique solution y(t) = S(t,t0 )y0 .

1.2.1 Well-Posedness From now on, we shall consider the value of t0 in the Cauchy problem (1.9) as fixed; in fact, unless otherwise specified, we shall always choose t0 = 0. We are then interested in the question of the dependence of solutions of (1.9) on the other initial value u0 . This question is naturally related to the WELL - POSEDNESS of the Cauchy problem (1.9). This means that solutions of (1.7) should not only be uniquely determined by the choice of the initial value u0 , but they should also depend continuously on u0 , in a specified topology. Since we are interested in the long-time behavior of the solutions, a crucial distinction has to be made between the notion of well-posedness on arbitrary, but bounded, time intervals [0, T ], and that of well-posedness in the whole interval [0, +∞[. Explicitly, we explain the first of these notions in DEFINITION 1.5 The Cauchy problem (1.9) is WELL POSED IN THE LARGE if for all u0 ∈ X , and all T , ε > 0, there exists δ > 0 such that for all v0 ∈ X and all t ∈ [0, T ], ku0 − v0 k < δ

=⇒ ku(t) − v(t)k < ε ,

(1.13)

where u and v are the unique solutions of (1.7) with u(0) = u0 and v(0) = v0 , and k · k denotes the norm of X . We remark that, in the theory of finite dimensional dynamical systems, definition 1.5 is often referred to as “continuity with respect to time and initial conditions”. Note that, in (1.13), δ depends not only on the initial value u0 , but, in general, also on T . That is, we can define a function (ε, T ) 7→ δ (ε, T ) (this function may often be defined only implicitly). If δ can be chosen independently of T , the solutions of (1.7) depend continuously on the initial data on all of [0, +∞[; this corresponds to the Lyapunov stability of the solutions of (1.7) (see definition A.6). In contrast, it

1.2

7

Ordinary Differential Equations

is well known that well-posedness in the large is not sufficient to guarantee stability, since the dependence of δ on T may be “bad”, in the sense that lim inf δ (ε, T ) = 0 .

T →+∞

To show this, it is sufficient to consider the following elementary example. Example 1.6 Consider the Cauchy problems for the ODEs u˙ = −u , u˙ = u ,

(1.14) (1.15)

with initial data at t = 0. Both problems have globally defined unique solutions for each choice of initial values, but only the first is globally well posed for t ≥ 0. In fact, when checking (1.13) we can take δ = ε for (1.14), but for (1.15) we are forced to take δ = εe−T , so in this case δ → 0 as T → +∞. We can interpret this in another way, realizing that the effect of any error in the initial value for equation (1.14) becomes negligible, up to arbitrary tolerance, if sufficient time is allowed to pass; on the contrary, even if two solutions of equation (1.15) are initially very close, after sufficient time they will be arbitrarily apart. Indeed, for (1.14), given any M and ε > 0, even if initially |u0 − v0 | ≥ M, it will be |u(t) − v(t)| ≤ ε for all t ≥ ln(M/ε), while for (1.15), given again any M and ε > 0, even if initially |u0 − v0 | ≤ ε, it will be |u(t) − v(t)| ≥ M for all t ≥ ln(M/ε). For instance, if we approximate u0 = π by v0 = 3.141, the initial error is less than 10−3 , but for the corresponding solutions of (1.15) we have |u(t) − v(t)| ≥ 103 for all t ≥ ln(103 /(π − 3.141)) ≈ 14.5087. This phenomenon is illustrated in figures 1.1 and 1.2. In terms of Lyapunov stability, the u large difference

small difference

0 large t

t

Figure 1.1: Exponential stability for u˙ = −u: A large difference in initial values still results in a small difference of the solutions after sufficient time.

8

1

Dynamical Processes

u large difference

small difference

0

large t

t

Figure 1.2: Exponential loss of information for u˙ = u: Even a small difference in initial values is drastically amplified after sufficient time.

point u = 0 in the phase space X = R, which corresponds to the solution u(t) ≡ 0 of both equations, is (uniformly) stable only for system (1.14), while system (1.15) is highly unstable under arbitrarily small perturbations of the initial value u0 = 0. In fact, if in (1.15) ±u0 > 0, then as t → +∞, u(t) → ±∞ (exponentially, of course), even if |u0 | < ε. Loosely speaking, this means that all control on the solution is lost if sufficient time is allowed to elapse.

1.2.2 Regular and Chaotic Systems As we have mentioned, the theory of dynamical systems is largely concerned with the behavior of the orbits t 7→ u(t) as t → +∞, and, more specifically, with how such behavior is influenced by the choice of the initial value u0 . This explains the use of notations like (1.19) below, which emphasize the dependence of the solution, at each time t, on its initial value u0 . With a great degree of simplification, we distinguish between two kinds of situations, which we call REGULAR and CHAOTIC. This choice of terms is rather arbitrary, and by no means universal; indeed, we find many different definitions of regularity and chaos in the literature, and even among those definitions that are mathematically rigorous, no one is universally accepted. Rather, different definitions are preferred for different applications. Roughly speaking, regular systems are those for which perturbations in the initial values will influence the orbits only for a short period of time (called TRANSIENT). After this time, different orbits would have the same qualitative behavior, and in particular the same asymptotic behavior. This type of situation is usually described by theorems like those on the asymptotic stability of a system, or the existence of limit cycles. In some cases, the asymptotic behavior is even independent of the initial values, in the sense that two orbits, even if starting from two points that are arbitrarily apart, after sufficient time (i.e. the transient, whose length depends on how far apart the initial values are) they will be and remain arbitrarily close to each other,

1.2

Ordinary Differential Equations

9

and so exhibit the same qualitative asymptotic behavior. Chaotic systems are instead those for which a sort of opposite situation holds. That is, these systems are extremely sensitive to even small variations of the initial values, in the sense that “close” initial conditions eventually move arbitrarily apart. The evolution of this type of system will be “regular” for a short time only (this is in general a consequence of some result analogous to the well-posedness of ODEs on compact time intervals). However, if observed for sufficiently long time periods, these systems not only do not exhibit any indication of convergence towards any sort of stable or periodic configuration, but their evolution seems to be totally unpredictable. More precisely, we give the following DEFINITION 1.7 Let S = (S(t,t0 ))(t,t0 )∈T∗2 be a dynamical process on X . S is said to DEPEND SENSITIVELY ON ITS INITIAL CONDITIONS if there is R > 0 such that for all t0 ∈ T , x0 ∈ X , and all δ > 0, there are y0 ∈ X and t1 ∈ T , t1 ≥ t0 , such that 0 < kx0 − y0 k ≤ δ and kS(t1 ,t0 )x0 − S(t1 ,t0 )y0 k ≥ R .

(1.16)

We remark that the notion of sensitive dependence of a dynamical process on its initial conditions is a natural generalization of that of uniform Lyapunov instability for ODEs (see section A.1). Example 1.8 Consider the dynamical processes S1 and S2 in R generated, respectively, by the ODEs (1.14) and (1.15) of example 1.6. Then S1 is regular, while S2 depends sensitively on its initial conditions. In fact, given ε and R such that 0 < ε < R, any two solutions x and y of (1.14) which initially differ by R will be such that |x(t) − y(t)| ≤ ε for all t ≥ t0 := ln Rε . That is, t0 is the transient after which these solutions will always differ by at most ε. In contrast, any two solutions x and y of (1.15) which differ initially by ε will be such that |x(t) − y(t)| ≥ R for all t ≥ t0 := ln Rε (compare to (1.16)). Example 1.8 shows that there are dynamical processes for which no matter how close two initial values may be, if sufficient time is allowed to pass the corresponding orbits will be arbitrarily apart. That is, the asymptotic behavior of these systems, which is still completely and uniquely determined by their initial values (the systems are deterministic), may be drastically different. To put this in another way, in this type of system all relevant information carried by the initial data is rapidly lost, and, consequently, it becomes impossible to maintain any reasonable control on the evolution of the system. Examples of this kind of situation are the smoke of a cigarette, the dynamics of large populations, of traffic patterns, economic cycles, etc. Probably, the most familiar example is that of the various meteorological models for the evolution of weather, whose prediction is in general relatively accurate only in a short time range (and the shorter the time interval, the better the prediction), but after sufficient time all predictions lose any practical value. In section 1.4 we shall see some other simple examples of systems that exhibit

10

1

Dynamical Processes

chaotic behavior, as described by their being sensitive to their initial conditions. Before proceeding, we mention another possible way of describing chaotic systems, whereby “an orbit can begin roughly anywhere and end up roughly anywhere”. More precisely, given any two open subsets U and V of X , there is x0 ∈ U such that the corresponding orbit intersect V. For an exhaustive discussion of these, and other, possible descriptions of the chaotic behavior of a dynamical system, we refer e.g. to Robinson, [Rob99], and to Alligood, Sauer and Yorke, [ASY96]. Of course, the possibility of determining whether a given system is regular or chaotic (we should rather say, whether the system may exhibit chaotic features or is guaranteed not to) is of extreme importance in applications, for at least two reasons. First, actual initial values depend on physical measurements, and are therefore never “exact” (this is not just a problem of √ the “real world”: Even in a simple numerical exercise in ODEs, initial values like 2 can only be introduced within approximations). Second, because in practice we cannot afford to observe the evolution of a system for very long time periods (deadlines have to be met, computer simulation time is expensive . . . ). Moreover, even if we could, we are still bound to observations in finite time intervals, and there is no guarantee that any such period of time, in which we may see “irregular” behavior, is still not part of a very long transient, after which the system may yet settle into a regular evolution. In these notes, we are concerned with a sort of intermediate situation between the two extremes described above. There are in fact examples of systems, whose evolution may appear to be chaotic, and yet after sufficient time their solutions seem to settle into a pattern that preserves a certain degree of order, which allows for some control of the disturbances typical of a chaotic regime. This type of behavior is usually better seen in the state space X , to which the solution curves (u(t))t ≥0 belong. More precisely, these systems are characterized by the existence of some subsets of X , to which the solution curves appear to converge (in the topology of the phase space), as t → +∞. These subsets are therefore called ATTRACTING SETS, and can be thought of as a generalization of the sets, such as stationary points or limit cycles, that are known to be attracting for regular systems of ODEs. Thus, for example, if a bounded attractor exists, two solutions which started at close initial values may still be quite apart at arbitrary later times (indication of chaos), but their distance cannot be arbitrarily large, since they both converge to the same attractor. In this sense the system is still controllable. Thus, even if we cannot decide whether a given system is regular, it is clearly desirable that we be able to determine if it at least possesses an attractor. Indeed, if this is the case, we would then know that, even if the system may possibly evolve chaotically, it will nevertheless settle into some type of controlled behavior. This is of course of fundamental importance in applications.

1.2.3 Dependence on Parameters In many physical examples, the equation (1.7) which models the evolution of a dynamical process may also depend on various numerical parameters, such as, for instance, the dielectric and permeability constants in Maxwell’s equations, or the viscosity coefficient in Navier-Stokes’ equations of fluid dynamics. In this case, equa-

1.2

Ordinary Differential Equations

11

tion (1.7) takes the more general form u˙ = F(λ ,t, u) ,

λ ∈ Λ ⊂ Rm ,

(1.17)

and the corresponding solution operator also depends on the parameters λ . In applications, it is of course of great importance to have a good knowledge of how the evolution of a system is influenced not only by (small) variations of the initial value u0 , but also by (small) variations of these parameters. For example, if the arm of a robot has the task of repeatedly moving an object from one position to another, and its motion is governed by a differential equation like (1.17), we are interested in the choice of parameters that make such motion as smooth as possible, and to avoid those that may make it irregular or, worse, chaotic. We will not present any theoretical results on the dependence of dynamical systems, in particular infinite dimensional ones, on numerical parameters, since this topic is too extensive and specialized, and a large quantity of the available insights and results are most often obtained by means of extensive and robust numerical simulation. Indeed, an experimental analysis of the equations modelling many physical examples indicates that various kinds of bifurcation phenomena typically occur at different, increasing values of λ . We refer to Temam, [Tem88, ch. 1], for a very general outline of various scenarios that are possible.

1.2.4 Autonomous Equations 1. Most classical results on the theory of the asymptotic behavior of dynamical systems involve systems generated by evolution equations (1.7) that are AUTONOMOUS. These systems, which occur quite frequently in applications, correspond to the case when the function F in (1.7) is independent of t, that is, when (1.7) has the form u˙ = F(u) ,

(1.18)

with F : X → X continuous. For example, the heat equation (1.8) is autonomous. In this case, we can always reduce ourselves, by a shift of the time coordinate, to a fixed choice of t0 . This means that the operators of S have the form ˜ − τ) , S(t, τ) = S(t − τ,t0 ) =: S(t ˜ t ≥t is a one-parameter family of operators, i.e. a SEMIFLOW, where now S˜ = (S(t)) 0 on X . In particular, we choose t0 = 0 for simplicity. We use again the letter S to denote this one-parameter family; that is, we write S = (S(t))t ≥0 , and (1.11) reads S(t)u0 = u(t, 0, u0 ) . In particular, conditions (1.2) and (1.3) of definition 1.1 are satisfied if S is a GROUP of (not necessarily linear) operators on X , i.e. if S(0) = IX

(1.19) SEMI -

(1.20)

12

1

Dynamical Processes

(the identity in X ), and for all t, s ≥ 0, S(t + s) = S(t)S(s)

(1.21)

(fig. 1.3). Indeed, if S is the solution operator defined by the autonomous equation

S(t1 )u S(t1 + t2 )u u

t2

t1 + t2

t1

Figure 1.3: The action of the semigroup.

(1.18), (1.20) holds by virtue of the initial condition S(0)u0 = u0

for all u0 ∈ X .

To show (1.21), we note that for all t, s ≥ 0, S(t)S(s)u0 = v(t) , where v is the solution of the Cauchy problem ( v˙ = F(v) , v(0) = S(s)u0 = u(s) . Thus, setting w(t) := v(t − s), we have w(t) ˙ = v(t ˙ − s) = F(v(t − s)) = F(w(t)) and w(s) = v(0) = u(s) . By the assumed uniqueness of solutions of the differential equation, we conclude that w(t) = u(t) for all t ≥ 0. In particular, u(t + s) = w(t + s) = v(t) ; and since u(t + s) = S(t + s)u0 and v(t) = S(t)v(0) = S(t)S(s)u0 ,

(1.22)

1.2

Ordinary Differential Equations

13

(1.22) means that (1.21) holds. Clearly, this argument may fail if the differential system is not autonomous, since then v(t ˙ − s) = F(t − s, v(t − s)) , and in general F(t − s, w) 6= F(t, w) . Example 1.9 The first order autonomous system (

u˙ = v v˙ = −u

(1.23)

generates the dynamical system S in R2 , defined by S(t)(x, y) := A(t)(x, y)> ,

t ∈ R,

where A(t) is the 2 × 2 matrix defined by A(t) :=

cost − sint

sint cost

! ,

and > denotes transposition. Indeed, it is immediate to verify that for all (x, y) ∈ R2 , the vector function t 7→ U(t) := A(t)(x, y)> solves system (1.23) with initial values U(0) = (x, y)> . Furthermore, A(0) = I, and for all t and s ∈ R, ! cos(t + s) sin(t + s) A(t + s) = − sin(t + s) cos(t + s) ! cost cos s − sint sin s sint cos s + cost sin s = − sint cos s − cost sin s cost cos s − sint sin s ! ! cos s sin s cost sint = − sin s cos s − sint cost = A(t)A(s) .

Example 1.10 The solution operator defined by the ODE u˙ = cost

14

1

Dynamical Processes

is not a semigroup. Indeed, for arbitrary t and s ∈ R we have u(t) = u0 + sint , S(t + s)u0 = u0 + sin(t + s) , S(t)S(s)u0 = u0 + sin s + sint . On the other hand, the solution operator defined by the autonomous ODE u˙ = 1 − u is indeed a semigroup. In fact, for arbitrary t and s ∈ R we have u(t) = (u0 − 1)e−t + 1 , S(t + s)u0 = (u0 − 1)e−(t+s) + 1 , S(t)S(s)u0 = (S(s)u0 − 1) e−t + 1 = (u0 − 1)e−s + 1 − 1 e−t + 1 = (u0 − 1)e−s−t + 1 .

Except for some elementary introductory examples, in these lectures we shall only consider autonomous systems. For an extensive account on the nonautonomous case, see e.g. Haraux, [Har91]. 2. When a system is autonomous, we call the corresponding family S of solution operators a SEMIFLOW on X , and the space X is often called the PHASE SPACE of the dynamical system. The map u : [0, +∞[ → X defined by u(t) := S(t)u0 ,

t ≥ 0,

is called a MOTION, and the image of u in X , i.e. the subset (or curve) γu0 :=

[

u(t) ⊂ X ,

t ≥0

is called the ORBIT of the motion u, starting at u0 . (When the system is not autonomous, we would need to consider the product R × X as an extended phase space.) Then, the asymptotic behavior of solutions of (1.18) is related to the evolution of the corresponding orbits, as subsets of X . Indeed, the recourse to the notion of orbits in the phase space (as opposed to that of solution of the differential equation) quite naturally allows us to introduce, together with the appropriate instruments from analysis in metric spaces to determine limiting behaviors etc., a more geometric approach, in which we study and exploit the topological properties of the orbits, seen in their own right as subsets of the phase space X . The example of definition of stability in the theory of ODEs is a familiar one; another example in two dimensions of space, i.e. for X = R2 , is the Poincaré-Bendixon theorem (see e.g. theorem A.32), which describes conditions under which the orbits of an autonomous system of two ODEs converge, in a suitable sense, to a limit cycle.

1.2

Ordinary Differential Equations

15

3. In conclusion, we have seen in what sense an autonomous differential equation (1.18) generates a continuous semiflow S, by means of the solution operator defined in (1.19). If F is sufficiently regular, S is also differentiable. It is worth to point out that the converse is also true; that is, a differentiable semiflow S = (S(t))t ≥0 is always generated by an autonomous ODE. PROPOSITION 1.11 Let S be a semiflow defined on X , and assume that for all x0 ∈ X , the map [0, +∞[ 3 t 7→ S(t)x0 ∈ X is differentiable at t = 0. Let F : X → X be defined by d F(x) := (S(t)x) , x ∈ X , dt t=0 and, for x0 ∈ X and t ≥ 0, set x(t) := S(t)x0 . Then x is differentiable in [0, +∞[, and satisfies the autonomous Cauchy problem ( x˙ = F(x) , (1.24) x(0) = x0 . PROOF Fix t0 ≥ 0. For t ≥ t0 , we compute that x(t) − x(t0 ) S(t)x0 − S(t0 )x0 S(t − t0 + t0 )x0 − S(t0 )x0 = = t − t0 t − t0 t − t0 S(t − t0 )S(t0 )x0 − S(t0 )x0 . = t − t0 Let y0 := S(t0 )x0 and θ := t − t0 . Then from (1.25) x(t) − x(t0 ) S(θ )y0 − y0 = . t − t0 θ Since the map t 7→ S(t)y0 is differentiable at t = 0, we have that, as θ → 0 x(t) − x(t0 ) d −→ (S(θ )y0 ) = F(y0 ) = F(S(t0 )x0 ) . t − t0 dθ θ =0 This proves that x is differentiable from the right, and 0 (t) = F(x(t)) . x+

If instead 0 < t < t0 , we compute that 0 x− (t0 ) = lim

t →t0−

x(2t0 − s) − x(t0 ) x(t) − x(t0 ) = lim t − t0 t0 − s s→t0+

(1.25)

16

1 = lim

s→t0+

Dynamical Processes

S(t0 − s)y0 − y0 S(θ )y0 − y0 = lim t0 − s θ θ →0+

= F(y0 ) = F(x(t0 )) . This proves that x is differentiable also from the left, and 0 x− (t) = F(x(t)) .

Hence, x is differentiable, and satisfies the equation of (1.24). The initial value of (1.24) is obviously taken.

1.3 Attracting Sets We have mentioned that in some cases, even if the evolution of a system appears to be chaotic, a certain degree of order seems to be preserved, in the sense that the orbits of the system appear to settle into a somewhat regular pattern, described by the fact that they converge, or at least remain “close”, to some subset of X . We can often describe this situation in terms of subsets that are ATTRACTING, or at least ABSORBING , in the following sense. 1. Absorbing Sets. In the theory of ODEs, a first step in the study of the asymptotic behavior of the solution of a given system is to recognize that these solutions are bounded as t → +∞. Analogously, given a dynamical system S on a Banach space X , it may be possible, in some cases, to recognize the existence of a subset B ⊂ X into which all orbits, or at least those starting from some subset U ⊆ X containing B, enter and, after possibly leaving B a finite number of times, eventually remain in B for ever. This set B is thus called an ABSORBING SET. If a bounded absorbing set exists, this is taken as an expression of a specific property of the system, generically called DISSIPATIVITY. 2. Attracting Sets. When an absorbing set exists, it is sometimes possible to also recognize the existence of a smaller subset A ⊂ B, to which all orbits starting from U converge as t → +∞ after having entered B; see fig. 1.4. (The precise definition of convergence of an orbit to a set of X is given in section 2.1 of chapter 2.) Such sets A are generally called ATTRACTING SETS. We will see that if a dynamical system admits an attractor, it necessarily has an absorbing set as well. Attracting sets may have a quite complicated geometric or topological structure (they may be self-similar sets, or FRACTALS), and the convergence of the orbits to these sets may be quite slow. However, these sets often possess some important properties, that may allow for a better understanding of the evolution of the system (in particular, if the system appears to be chaotic). For example, the set A may be compact, and (often but not always) it may have a finite fractal dimension

1.3

17

Attracting Sets

A B

x

Figure 1.4: Absorbing and attracting sets. After entering the absorbing set B for the last time at x, the orbit remains in B, and then converges to A. (the definition of which we recall in section 2.8 of chapter 2). The set A may also be invariant, which means that S(t)A = A

for all t ≥ 0 .

(1.26)

That is, if u0 ∈ A then u(t) = S(t)u0 ∈ A for all t ≥ 0 and, conversely, every u0 ∈ A is on some orbit starting from some point in A. 3. Attractors. Bounded, positively invariant attracting sets are generally called ATTRACTORS . Of particular importance are attractors that are finite dimensional, because the corresponding dynamics is also finite dimensional. Indeed, the invariance of the attractor implies, by (1.26), that orbits which originate in the attractor remain in the attractor for all future times; consequently, the evolution of a system on a finite dimensional attractor would essentially be governed by a finite system of ODEs. In fact, a celebrated theorem of Mañé, [Mañ81], states that if a dynamical system possesses a finite dimensional attractor, this attractor can be generated by (or, as it is sometimes said, is “imbedded into”) a finite system of ODEs. This result allows us to reduce, at least in principle, the study of the long time behavior of orbits which converge to a finite dimensional attractor to that of the solutions of a finite dimensional system of ODEs on A. This question, together with the description of the corresponding ODEs, is one of the most challenging problems in the theory of dynamical systems. Moreover, in most cases the reduction of the study of the evolution of the system on the attractor cannot be pursued in practice, because of several difficulties, which partially motivate the search for “friendlier” sets, such as the inertial manifolds discussed below. For example, we have mentioned the generally nonsmooth geometrical or topological structure of the attractor, which may cause the corresponding ODEs to only admit generalized solutions. Another problem, of special importance in applications, is that in many cases the available estimates on the dimension of the attractor, and therefore on the dimension of the system of ODEs, are simply too large for computational feasibility. For instance, in meteorology it is not uncommon to have estimates of the

18

1

Dynamical Processes

order of 10m , m ≥ 20. Also, attractors are in most cases not sufficiently stable under perturbations of the data, so that their numerical approximations, and the consequent propagation of errors, may be quite difficult to control. For example, approximations of attractors with respect to the Hausdorff distance (see section 2.1) are in general only upper semicontinuous. Finally, the rate of convergence of the orbits to the attractor may really be no better than polynomial, as the following example shows. Example 1.12 Consider the semiflow S generated by the autonomous ODE u˙ = f (u) := −u3 .

(1.27)

The attractor of S is the set A = {0}, but the convergence of the orbits to A is at most polynomial, as we see from the explicit solution of the Cauchy problem relative to (1.27) with initial value u(0) = u0 , that is, u0 u(t) = q . 1 + 2u20t

4. Inertial Manifolds. On the other hand, there are systems whose attractors do not present this type of difficulties, since they are imbedded into a finite dimensional Lipschitz manifold M of X , and the orbits converge to this manifold with a uniform exponential rate. Such a set M is called an INERTIAL MANIFOLD of the system (fig. 1.5). When an inertial manifold exists, the evolution of the semiflow on the

Figure 1.5: Inertial Manifolds.

manifold is governed by a finite system of ODEs, called the INERTIAL FORM of the semiflow. This finite system of ODEs will in general admit solutions with a certain degree of smoothness, depending on the smoothness of the manifold. Since orbits converge to the inertial manifold with a uniform exponential rate, we see that, in turn, the dynamics on the manifold will be a good description of the long time behavior of solutions of equation (1.7). Clearly, the possibility of imbedding the attractor into an

1.3

Attracting Sets

19

inertial manifold provides an indirect way to obtain the above mentioned desired system of ODEs. Moreover, the uniformity of the rate of convergence of the orbits to the manifold makes these systems extremely stable under perturbations and numerical approximations. Unfortunately, there are not many examples of systems which are known to admit an inertial manifold; among these, we mention the semiflows generated by a number of reaction-diffusion equations of “parabolic” type, and by the corresponding hyperbolic (small) perturbations of these equations. A typical model is that of the so-called Chafee-Infante equations, which we present in chapter 5. 5. Exponential Attractors. An intermediate situation occurs when a system admits a so-called EXPONENTIAL ATTRACTOR. These sets, which are also sometimes called INERTIAL SETS in the literature, are somehow intermediate between attractors and inertial manifolds, in the sense that while they do not necessarily have a smooth structure, they can still be imbedded into a finite system of ODEs. In addition, these sets retain at least three of the features of the inertial manifolds that attractors do not necessarily have: the finite dimensionality, the exponential convergence of the orbits, and a high degree of stability with respect to approximations (for example, continuity with respect to the Hausdorff distance). This means that when an exponential attractor exists, after an “exponentially short” transient the dynamics of the system are essentially governed by a finite system of ODEs (the classical image is that of an airplane, landing at a “fast” speed and then “slowly” taxiing to the arrival gate). The following is a simple example of a regular system, whose solutions converge exponentially to its attractor. Example 1.13 Consider the function f : [0, 1] → [0, 1] defined by f (x) = (1 + x)−1 , and the corresponding discrete system (Sn )n∈N defined by the iterated sequence (1.30). This √ system has an attractor, which is the set A = {`}, with ` := ( 5 − 1)/2. We now show that A is also an exponential attractor; that is, there is α > 0 such that, for all initial values x0 ∈ [0, 1], |Sn x0 − `| ≤ e−αn .

(1.28)

Indeed, setting Sn x0 = f (xn ) =: xn+1 , we see that this sequence converges to the positive solution of the equation x = f (x), which is precisely `. Since ` = f (`), we compute that ` − xn+1 = f (`) − f (xn ) = Since 1 + ` >

3 2

xn − ` . (1 + `)(1 + xn )

and 1 + xn ≥ 1 for each n, we deduce from (1.29) that |xn+1 − `| ≤ 32 |xn − `| ,

(1.29)

20

1

Dynamical Processes

from which we conclude that |xn − `| ≤ ( 32 )n |x0 − `| ≤ ( 23 )n . This shows that (1.28) holds, with e.g. α = ln 32 > 0. We explicitly note that α is independent of the initial values: this ensures that the iterates Sn x0 converge to A with a uniform rate. Of course, not all dynamical systems will possess attractors, exponential attractors or inertial manifolds. In the sequel, we shall try to present a theory, by now quite well established, that provides a number of sufficient conditions on the system for at least some of these sets to exist. In particular, since attractors will contain stationary and periodic solutions of (1.17), this theory is really a natural extension of the classical theory of stability for ODEs.

1.4 Iterated Sequences Not surprisingly, many of the ideas (and difficulties) in the theory of continuous dynamical systems already surface in the context of discrete dynamical systems generated by ITERATED SEQUENCES. These are sequences (un )n∈N ⊂ X , of the form un+1 = f (un ) ,

(1.30)

where f is a map of X into itself. Thus, each sequence is completely determined by its initial value u0 , assigned separately. Iterated sequences generate a DISCRETE dynamical system S := (Sn )n∈N on X , defined by S0 = I , Sn+1 = S ◦ Sn , where Sn is the n-th iterate of S, and ◦ denotes the composition of maps in X . Thus, T = N, and the orbits of S are the sequences (Sn u0 )n∈N . We are interested in how the behavior of each such sequence, as n → +∞, depends on its initial term u0 . In this section we present some well known examples of discrete systems in Rn , n ≤ 3, each defined by a sequence like (1.30). For future reference, we recall the following DEFINITION 1.14

Let F : X → X be a map (not necessarily linear), and x ∈ X .

1. x is a FIXED POINT of F if x = F(x). 2. A fixed point x of F is said to be STABLE if, given any neighborhood U of x there ˜ the corresponding is another neighborhood U˜ ⊂ U of x such that for all x0 in U, recursive sequence (xn )n∈N , starting at x0 and defined by xn+1 = F(xn ), is contained in U. Otherwise, x is said to be UNSTABLE.

1.4

21

Iterated Sequences

3. A fixed point x of F is said to be ATTRACTIVE if for all x0 in a neighborhood of x, the above defined recursive sequence (xn )n∈N converges to x. 4. A stable and attractive fixed-point is called

ASYMPTOTICALLY STABLE .

5. A point x is said to be p- PERIODIC (p ∈ N) if F p (x) = x. Note that not all stable fixed points are attractive, as we see by taking F(x) = x. For this map, each point x is a stable, but not attractive, fixed point. On the other hand, we have the following THEOREM 1.15 Let X = R, and x0 be a fixed point of a C1 function F. Then x0 is asymptotically stable if |F 0 (x0 )| < 1, while if |F 0 (x0 )| > 1, x0 is unstable. PROOF Without loss of generality, we can confine ourselves to symmetric neighborhoods of x0 . 1) Assume first that |F 0 (x0 )| < 1. There exists then a number ε ∈ ]|F 0 (x0 )|, 1[, and, correspondingly, a number δ > 0 such that if |x − x0 | ≤ δ , then |F(x) − F(x0 )| ≤ ε|x − x0 | .

(1.31)

Since ε < 1 and F(x0 ) = x0 , (1.31) implies that |F(x) − x0 | ≤ |x − x0 | ≤ δ . Consequently, we can repeat estimate (1.31), and obtain that for all n ∈ N≥1 , |F n (x) − x0 | ≤ ε n |x − x0 | .

(1.32)

From this, it follows that x0 is asymptotically stable: Indeed, given any neighborhood U := ]x0 − ρ, x0 + ρ[ of x0 , let δ ∈ ]0, ρ], and set U˜ := ]x0 − δ , x0 + δ [. Then, (1.32) ˜ each iterate F n (x) is in U, because implies that if x ∈ U, |F n (x) − x0 | ≤ ε n |x − x0 | ≤ δ ≤ ρ . Thus, x0 is stable; clearly, (1.32) also implies that x0 is also attractive. 2) Conversely, assume that |F 0 (x0 )| > 1. Then, as before, given any a ∈]1, |F 0 (x0 )|[, we can determine γ > 0 such that if |x − x0 | ≤ γ, then |F(x) − x0 | ≥ a|x − x0 | .

(1.33)

¯ there are x¯ and n¯ such We wish to prove that there is ρ¯ > 0 such that for all δ ∈]0, ρ], that |x¯ − x0 | ≤ δ

and

|F n (x) ¯ − x0 | ≥ ρ¯ .

22

1

Dynamical Processes

Arguing by contradiction, taking ρ = γ, we can determine δ ∈ ]0, γ] such that if |x − x0 | ≤ δ , then for all n ∈ N>0 , |F n (x) − x0 | ≤ ρ = γ .

(1.34)

Now, (1.34) and (1.33) imply that for all n, |F n (x) − x0 | ≥ an−1 |F(x) − x0 | ;

(1.35)

but since |x − x0 | ≤ δ ≤ γ, (1.33) implies that, in fact, |F n (x) − x0 | ≥ an |x − x0 |

(1.36)

for all n. Choose then, for example, x = x0 + 21 δ . Then, (1.36) implies that γ ≥ |F n (x) − x0 | ≥ 21 δ an .

(1.37)

Since a > 1, letting n → +∞ in (1.37) we achieve the desired contradiction.

x x x0 1x2 3

x3 x1 x2 x0

x1 x x0 3 x2

x1

x3 x2 x0

Figure 1.6: The four possibilities: F 0 (x0 ) > 1, F 0 (x0 ) < −1, 0 < F 0 (x0 ) < 1, −1 < F 0 (x0 ) < 0. We remark that when |F 0 (x0 )| = 1, x0 can be either attractive, or unstable. This is easily seen by considering a function F which changes concavity at x0 . For example, if F 0 (x0 ) = 1, and F changes from convex to concave at x0 , then x0 is attractive, while if F changes from concave to convex at x0 , then x0 is unstable.

1.4.1 Poincaré Maps Given a continuous dynamical system, it is in many cases possible to construct a discrete one, whose asymptotic behavior is essentially the same as that of the continuous system. One way to do so is to choose a sequence (tn )n∈N of equidistant values tn → ∞ and, given a solution of the continuous autonomous system (1.18), to consider the corresponding sequence (un )n∈N of points un := u(tn ) in the phase space X . Clearly, each of these points lies on the orbit starting at u0 . This choice defines a map Φ : X → X , by un+1 = Φ(un ) .

(1.38)

1.4

Iterated Sequences

23

Maps constructed in this way are called STROBOSCOPIC MAPS. For example, the choice tn = n + 1 in (1.38) yields the sequence (un )n∈N , defined by S := S(1) ,

un+1 = Sn u0

for n ∈ N .

We can visualize a stroboscopic map by considering the graph of u in the product space [0, +∞[ ×X ; that is, the set graph u := {(t, u(t)) : t ≥ 0} .

(1.39)

Then, the sequence in X defined by the stroboscopic map (1.38) is obtained by projecting on X the points (tn , u(tn )). In the case of finite dimensional systems, a remarkable construction is that of the so-called P OINCARÉ MAPS. These maps are constructed by fixing a hyperplane Σ ⊂ Rn , called a P OINCARÉ SECTION, and considering on Σ the sequence of points Pn defined by the “first returns” of the (graph of the) solution on Σ , i.e. by the successive intersections of the semiorbit {u(t) : t ≥ 0} with Σ (figs. 1.7 and 1.8). Indeed,

Σ

Figure 1.7: The Poincaré section.

Poincaré maps are sometimes also known as “first return” maps. More precisely, we consider again the intersection of the graph (1.39) with R × Σ (both as subsets of R × Rn ), and construct the sequence of points (u(tn ))n∈N ⊆ Σ , as ordered by the first argument tn ; that is, by the time of the n-th intersection of the orbit with the hyperplane Σ . Set un := u(tn ). The sequence (un )n∈N can then be considered as a recursive sequence on Σ , defined by a map un+1 = ΦΣ (un ) .

24

1

Dynamical Processes

Σ

limit cycle

Figure 1.8: In the plane, the Poincaré section is a line. The map ΦΣ is called the P OINCARÉ MAP associated to the semiflow defined by (1.18). Poincaré maps can thus be used to study the asymptotic behavior of a continuous semiflow, by reducing it to a discrete one. For example, if (1.7) has a periodic solution with period T , the Poincaré map with sampling synchronized with the period, i.e. with tn = nT , will have a fixed point (fig. 1.9). Of course, for a given ODE, or system of ODEs, even autonomous ones, it may not be clear how to find suitable sampling sequences (tn )n∈N , and extensive numerical experimentation may well be required. Finally, we mention that the notion of Poincaré maps can be generalized to infinite dimensional continuous dynamical systems (see e.g. Marsden-McCracken, [MM76]).

1.4.2 Bernoulli’s Sequences We start with an example that illustrates the phenomenon of the loss of information from the initial data after sufficient time is allowed to pass. The so-called B ERNOULLI ’ S SEQUENCE is the recursive sequence xn+1 = f (xn ) generated by the function f : [0, 1] → [0, 1] defined by f (x) := 2x − b2xc , where bxc denotes the integer part of x (that is, the largest integer less than or equal to x). Note that f is not continuous at x = 12 (fig. 1.10); however, f can be nicely described as a so-called “circle-doubling” map, if we identify the endpoints of the domain interval [0, 1] with each other. More precisely, if we define g : [0, 1] → R2 by g(x) := (cos(2π f (x)), sin(2π f (x))) ,

1.4

25

Iterated Sequences

periodic orbit Figure 1.9: Periodic and 2-periodic orbits produce fixed points in a Poincaré section. 1

1 2

1

Figure 1.10: f (x) = 2x − b2xc then g is continuous also at x = 21 . In fact, the point g(x) runs twice around the unit circle, once as x ∈ [0, 21 [, and another as x ∈ [ 12 , 1[. It is easy to study the stability of Bernoulli’s sequences: x = 0 is the only stationary point of f , and if x0 = 12 or x0 = 1, then x1 = 0, so xn = 0 for all n ≥ 1. Consider then any other initial value x0 different from 0, 21 and 1, and let ε > 0 be such that y0 := x0 + ε is in the same half interval ]0, 12 [ or ] 21 , 1[ which contains x0 . Then y1 − x1 = f (x0 + ε) − f (x0 ) = 2ε . Next, if y1 and x1 are both still in the same half interval, we proceed to compute in the same way that y2 − x2 = f (y1 ) − f (x1 ) = 2(y1 − x1 ) = 4ε . Proceeding in this fashion, we see that yn − xn = 2n ε, as long as yn−1 and xn−1 are in the same half interval. This computation shows that the distance between orbits

26

1

Dynamical Processes

grows exponentially; this has the consequence that at a certain point the orbits “must separate”, no matter how close they were initially. In fact, we have that yn − xn ≥ 12 1 as soon as n ≥ log2 2ε , and after this point the difference yn+1 − xn+1 is no longer controllable. One way to interpret this situation is that all information deriving from the knowledge of x0 is eventually lost. For example, if x0 represents the “true” initial value in an experiment, and x0 ± ε is its actual measurement, after a number of steps equal 1 no meaningful control of the error between the true and the approximated to log2 2ε initial values is maintained. Figure 1.11 illustrates this phenomenon, by comparing

y2

y0 x0

x1

y1

x2

Figure 1.11: Two Bernoulli sequences.

the evolution of the sequences corresponding to different approximations x0 and y0 of π4 . After the first iteration, the points x1 and y1 are still close, but at the second iteration, x2 ≈ 1, while y2 ≈ 0. In fact, the evolution of Bernoulli’s sequence is chaotic, in the sense that it is sensitive to its initial conditions (see definition 1.7). We actually show more than (1.16); namely, we show that there is R > 0 such that for all δ ∈ ]0, 1[ and all x0 , y0 ∈ [0, 1] such that |x0 − y0 | ≤ δ , there is k ∈ N such that |xk − yk | ≥ R. To see this, we proceed by contradiction. Thus, we assume that for all r > 0 there are δr ∈ ]0, 1[ and x0 , y0 ∈ [0, 1] such that |x0 − y0 | ≤ δr , and for all k ∈ N, |xk − yk | ≤ r .

(1.40)

We fix r = 41 , and determine δ1/4 , x0 and y0 accordingly. Our previous discussion implies that there are infinitely many indices m ∈ N such that xm and ym are in different

1.4

27

Iterated Sequences

half-intervals. In fact, if instead there were m0 such that the sequences (xm )m≥m0 and (ym )m≥m0 are in the same half-interval, then we would deduce that 1 2

≥ |xm − ym | = 2m−m0 |xm0 − ym0 | ,

which is a contradiction as m → +∞. It follows that there is at least one pair of subsequences (xm p ) p∈N and (ym p ) p∈N such that for all p ∈ N, either xm p <

1 2

≤ ym p

ym p <

or

1 2

≤ xm p .

(1.41)

If e.g. the first of (1.41) holds, recalling that |xm p − ym p | ≤ 41 we easily see that xm p ≥ ym p ; therefore, recalling(1.40), we deduce the contradiction 1 4

≥ |xm p +1 − ym p +1 | = xm p +1 − ym p +1 = 1 − 2 ym p − xm p ≥ 12 .

With a totally analogous computation, we find the same result if the second of (1.41) holds. Hence, we conclude that Bernoulli’s sequence is sensitive to its initial conditions. The loss of information characteristic of Bernoulli’s sequence can be described explicitly. Indeed, let x0 be represented in the binary system by the series ∞

x0 =

αn , n n=1 2

∑

αn ∈ {0, 1} .

Then ∞

αn x1 = 2x0 − b2x0 c = ∑ n−1 − 2 n=1

$

∞

αn ∑ 2n−1 n=1

%

∞

∞ αn αn+1 − α1 = ∑ n . n − 1 2 n=2 n=1 2

= α1 + ∑

This means that f moves the digits of the fractional part of each number xn one position to the left, and subtracts the unit that may so result. For example, if 1 1 x0 = 0.1101001 = 12 + 41 + 16 + 128 ,

then 1 x1 = (1 + 21 + 81 + 64 ) − 1 = 0.101001 .

Now, in any numerical approximation, the initial value x0 is known only up to a finite number of digits of its fractional part. If m is this number, after m iterations of Bernoulli’s map we obtain xm = 0; that is, we reach the fixed point of the map. Thus, all information from x0 is lost in a finite number of steps.

1.4.3 Tent Maps Another example of the phenomenon of the strong dependence of a system on its initial data, which numerically translates into a drastic loss of significant information,

28

1

Dynamical Processes

is provided by the iterated sequence (1.30), corresponding to the family of functions fλ : [0, 1] → [0, 1] defined for λ > 0 by ( 2λ x for 0 ≤ x ≤ 21 , fλ (x) = (1.42) 2λ (1 − x) for 21 < x ≤ 1 . Each fλ is an example of a so-called

1

TENT MAP

(fig. 1.12). Note that if λ = 21 , the

λ =1

λ=

1 2

1 2

Figure 1.12: Tent maps for λ =

1 1 2

and λ = 1.

sequence is constant at least after its second term; otherwise, it is possible to show (see e.g. Moon, [Moo92]), that the evolution of the corresponding dynamical system is regular if λ < 21 , and chaotic (in the sense that it is sensitive to its initial conditions, see definition 1.7) if λ > 12 . In particular, we show this for λ = 1. Example 1.16 Let S be the dynamical system in X = [0, 1] defined by the function f1 of (1.42), which can be written as f1 (x) = 1 − |1 − 2x| =: f (x) .

(1.43)

Then, S is sensitive to its initial conditions. To show this, given x0 ∈ [0, 1], we consider the corresponding recursive sequence defined by xn+1 = f (xn ) (i.e., the orbit of S starting at x0 ). As was the case for Bernoulli’s sequences, there are infinitely many terms of these sequences that fall in each of the subintervals L := [0, 21 ] and R := [ 12 , 1]. In fact, for each k ∈ N>0 , it is possible to keep track of the half interval in which the term xk will fall, by means of the following device, which is adapted from Alligood, Sauer and Yorke, [ASY96,

1.4

Iterated Sequences

29

ch. 1.8]. Given k choices S1 , . . . , Sk of the letters L or R, we define a subinterval S1 . . . Sk of [0, 1] , by x0 ∈ S1 . . . Sk

⇐⇒

x j ∈ S j+1

for j = 0, . . . , k − 1

(1.44)

(thus, S1 is either L or R). There are exactly 2k such subintervals, each having length 1 ; we can order these in a family I := {I1 , . . . , I2k }. Note that 2k [

I j = [0, 1] ,

(1.45)

1≤ j≤2k

and that, if x0 ∈ S1 . . . Sk , then x1 ∈ S2 . . . Sk , x2 ∈ S3 . . . Sk , . . . , xk−1 ∈ Sk .

(1.46)

The way the family I is actually ordered is of no importance here, except for the case k = 2, in which we have the ordering [0, 1] = LL ∪ LR ∪ RR ∪ RL = [0, 41 ] ∪ [ 41 , 12 ] ∪ [ 21 , 34 ] ∪ [ 43 , 1] ,

(1.47)

which is easily verified. By way of illustration, we consider the case k = 3, where the 8 subintervals are LLL = [0, 81 ] , LLR = [ 81 , 41 ] , LRR = [ 14 , 83 ] , LRL= [ 38 , 12 ] , RRL = [ 12 , 85 ] , RRR= [ 85 , 43 ] , RLR = [ 34 , 87 ] , RLL = [ 78 , 1] (in accord with (1.45)). For instance, suppose that x0 ∈ LRL. Then, by definition (1.44), x0 ∈ L = [0, 12 ] , x1 ∈ R = [ 12 , 1] , x2 ∈ L = [0, 12 ] ,

(1.48)

and this is in accord with (1.46). By (1.43), the first of (1.48) implies that x1 = 2x0 . The second of (1.48) implies then that 14 ≤ x0 ≤ 21 ; i.e., x0 ∈ LR. Again by (1.43), the second of (1.48) also implies that x2 = 2 − 2x1 = 2 − 4x0 ; then, the third of (1.48) finally implies that 38 ≤ x0 ≤ 12 , as claimed. We are now ready to show the sensitivity of the semiflow S to its initial conditions. We claim that, given δ ∈ ]0, 1[ and x0 ∈ [0, 1], there are y0 ∈ [0, 1], with 0 < |x0 − y0 | ≤ δ , and k ∈ N>0 , such that |xk − yk | ≥

1 4

(1.49)

(compare to (1.16)). To show this, fix δ ∈ ]0, 1[ and x0 ∈ [0, 1]. Let k ∈ N>0 be such that 21k < δ . Then the interval ]x0 − δ , x0 + δ [ ∩ [0, 1] contains at least one subinterval S1 . . . Sk Sk+1 Sk+2 , with x0 ∈ S1 . . . Sk Sk+1 Sk+2 . Fix one of these subintervals, for which there are the four possibilities S0 . . . Sk LL , S0 . . . Sk LR , S0 . . . Sk RR , S0 . . . Sk RL .

30

1

Dynamical Processes

We choose then y0 in the subinterval where x0 is, but with the last letters L and R reversed; that is, respectively, in S0 . . . Sk RR , S0 . . . Sk RL , S0 . . . Sk LL , S0 . . . Sk LR . Then, since both x0 and y0 are in the larger subinterval S1 . . . Sk , whose length is 1 < δ , we have that |x0 − y0 | < δ , and (1.49) holds. Indeed, suppose e.g. that 2k x0 ∈ S0 . . . Sk LL , y0 ∈ S0 . . . Sk RR . Then, by (1.46), xk ∈ LL and yk ∈ RR, which, recalling (1.47), means that 0 ≤ xk ≤ 14 ,

1 2

≤ yk ≤ 34 .

Thus, |xk − yk | = yk − xk ≥ 12 − 41 = 41 . An analogous argument shows that the same inequality holds in each of the remaining three possibilities. Thus, (1.49) follows, proving the sensitivity of the semiflow S defined by (1.43) to its initial conditions. From these two examples, we could surmise that the chaotic behavior of a discrete system may be a consequence of the fact that the functions that define the sequence (1.30) are not regular (for Bernoulli’s sequences, f is not continuous; for the tent maps, f is not differentiable). The next example shows that we can in fact have chaotic behavior even with C∞ maps.

1.4.4 Logistic Maps A regularized version of the tent maps is provided by the family of functions fλ : R → R defined by fλ (x) = λ x(1 − x) , see fig. 1.13. When 0 ≤ λ ≤ 4, each fλ maps the interval [0, 1] into itself; the corresponding iterated sequence (1.30) is called a LOGISTIC SEQUENCE. These sequences are a normalized version of the sequence (xn )n∈N , defined by xn+1 = axn − bxn2 , which describes a model of population growth; the coefficient a represents a constant growth rate, and b measures an external inhibiting factor. In absence of the latter, i.e. when b = 0, the terms of the sequence reduce to xn+1 = an+1 x0 . In this case, the fixed point x = 0 is stable if |a| < 1, and unstable if |a| > 1. If a = 1 we have a 1-periodic orbit, and if a = −1, a 2-periodic orbit.

1.4

x1

31

Iterated Sequences

x8

x4

x6

x2

x5

x3 x7

Figure 1.13: A sequence generated by the logistic map. To find the stationary points of the logistic sequence, we consider the equation f (x) = x (having dropped the index λ for convenience), i.e. x = λ x(1 − x) . This equation has the two solutions x = 0 and x = sλ := 1 − λ1 . Since sλ ∈ [0, 1] if and only if λ ≥ 1, we conclude that if λ < 1, x = 0 is the only stationary point. Since s1 = 0, the same is true for λ = 1. If x0 = 1, then xn = 0 for all n ≥ 1. Since f 0 (0) = λ , recalling theorem 1.15 we see that the stationary point x = 0 is stable if λ < 1, and unstable if λ > 1; similarly, since f 0 (sλ ) = 2 − λ , sλ is stable if 1 < λ < 3, unstable if λ > 3. We also see directly that x = s1 = 0 and x = s3 are stable also if, respectively, λ = 1 and λ = 3. As we know, stationary points of the sequence correspond to 1-periodic orbits. To find 2-periodic orbits, we look for the stationary points of the second iterate of f , i.e. for solutions of the equation x = f ( f (x)) =: f 2 (x) = λ 2 x(1 − x)(λ x2 − λ x + 1) , (fig. 1.14). Of course, f 2 (0) = 0 and f 2 (sλ ) = sλ , since a fixed point of f is also a fixed point of any of its iterates. Other fixed points of f 2 are found by solving the equivalent equation x − f 2 (x) =: Q(x) = 0 . x(x − sλ ) We compute that Q(x) = λ 2 λ x2 − (1 + λ )x + 1 + λ −1 ;

32

1

Dynamical Processes

f (x)

x f ( f (x))

x0

x1

x

Figure 1.14: 2-periodic orbits: x0 and x1 are fixed points of f ( f (x)). since the discriminant of Q equals ∆Q := (λ + 1)(λ − 3), for λ > 3 there are two more fixed points of f 2 ; these produce two 2-periodic orbits for the sequence. For λ = 3, ∆Q = 0, s3 = 23 , and Q(x) = −9(3x2 − 4x + 43 ) = −27(x − s3 )2 ; thus, there still is only one 1-periodic orbit. At this point, we should proceed with the study of the stability of these fixed points, and then look for fixed points of further iterates of f , and so on. However, the analytical complexity of these computations is such that it is far more effective to resort to numerical experimentation and geometric or topological arguments. As can be seen in Moon, [Moo92], the numerical evidence shows that, as λ increases to 4, the corresponding orbits first exhibit period doubling, then turn to chaotic behavior. In fact, proceeding as in example 1.16, we can easily show that the logistic sequence corresponding to λ = 4 (in which case the range of f4 is all of [0, 1]) is sensitive to its initial conditions.

1.5 Lorenz’ Equations In this and the next section, we present two examples of continuous, finite dimensional dynamical systems, which admit an attractor for certain values of their

1.5

Lorenz’ Equations

33

parameters. This values are found by numerical experiment; the existence of the attractor can be confirmed by the methods described in the next chapter.

1.5.1 The Differential System We begin with the so-called (and quite famous) L ORENZ ’ EQUATIONS , which are the system of the three autonomous differential equations    x˙ = −σ x + σ y (1.50) y˙ = rx − y − xz   z˙ = −bz + xy , with σ , r and b > 0. This system was proposed by Lorenz in [Lor63] as an approximation, with the three degrees of freedom σ , r, and b, of the Boussinesq equations modelling the convective motion of a stratified bidimensional fluid heated by convection from below, such as air over the earth’s surface. As such, it provides a model, admittedly oversimplified, of an atmospheric phenomenon of interest in meteorology. Our goal is to study the behavior of solutions to (1.50), in relation to the parameter r (the Rayleigh’s number), keeping σ and b fixed. In the next chapter we shall show that for all values of σ , b and r, and for all initial values (x0 , y0 , z0 ) ∈ X = R3 , system (1.50) has a unique global solution; that this system generates a semiflow S on R3 ; that S admits a bounded absorbing set in R3 ; that, as a consequence, system (1.50) has a compact attractor A. For certain values of r, the structure of this attractor is relatively well understood. Although most detailed information can be obtained by means of extensive numerical experimentation, we present here some results that can be established by simple analytical techniques. For a more extensive study of Lorenz’ equations, we refer e.g. Sparrow, [Spa82], Marsden-McCracken, [MM76], and Guckenheimer-Holmes, [GH83].

1.5.2 Equilibrium Points The stationary points of (1.50) are obtained by solving the system    −σ (x − y) = 0 rx − y − xz = 0   −bz + xy = 0 . We obtain that if r ≤ 1, the origin is the only equilibrium point, while if r > 1 there are exactly three equilibrium points: The origin O = (0, 0, 0) and the two other points p p C± := (± b(r − 1), ± b(r − 1), r − 1) . To study the stability of these three points, recalling theorem A.7 we linearize system (1.50) at each point, and consider the sign of the real part of the eigenvalues of the

34

1

Dynamical Processes

Jacobian matrix  −σ σ 0   J(x, y, z) =  r − z −1 −x  y x −b 

at each equilibrium point. At the origin the characteristic polynomial is  −(σ + λ ) σ 0   det[J(O) − λ I] = det  r −(1 + λ ) 0  0 0 −(b + λ ) 

= −(b + λ ) (λ 2 + (σ + 1)λ + σ (1 − r)) . | {z } := P(λ ) Thus, J(O) always has at least the real negative eigenvalue λ1 = −b. The discriminant ∆P of the polynomial P equals ∆P = (σ − 1)2 + 4σ r ; since ∆P > 0, J(O) has in fact three real eigenvalues. If r < 1, all coefficients of P are positive, so the eigenvalues are all negative, the unique equilibrium point O is a stable node, and is in fact the attractor of the system (i.e., A = {O}). If r > 1, one eigenvalue of J(O) is positive, so the origin is an unstable saddle, with a 2dimensional stable manifold Ms (O) attracted by O, and a one-dimensional unstable manifold Mu (O) repelled by it; see definition 2.22 in chapter 2. At the points C± , r > 1 and the characteristic polynomial is 

 −(σ + λ ) σ 0 p   det[J(C± ) − λ I] = det  1 −(1 + λ ) ∓ b(r − 1)  p p ± b(r − 1) ± b(r − 1) −(b + λ ) = −(λ 3 + (σ + b + 1)λ 2 + b(σ + r)λ + 2bσ (r − 1)) =: −P1 (λ ) . Again, at least one eigenvalue λ1 is real. The other two eigenvalues λ2 and λ3 are either both real, or complex conjugate. If they are real, elementary calculus shows that, since r > 1, both are negative; hence, C+ and C− are stable nodes. If instead λ2 and λ3 are complex nonreal, to study the sign of their real part we set λ2 = ζ = u + iv and λ3 = ζ = u − iv, and proceed as follows. Writing the characteristic equation as P1 (λ ) = (λ − λ1 )(λ − ζ )(λ − ζ¯ ) = 0 , we obtain the equation λ 3 − (2u + λ1 )λ 2 + (|ζ |2 + 2λ1 u)λ − λ1 |ζ |2 = 0 .

1.5

35

Lorenz’ Equations

We easily verify that, since v 6= 0, u = Re ζ = 0 if and only if the product of the coefficients of λ 2 and λ equals the constant term. In terms of the original form of P1 (λ ), this translates into the condition b(σ + b + 1)(σ + r) = 2bσ (r − 1) .

(1.51)

As an equation in r, if σ 6= b + 1, (1.51) has the solution r∗ =

σ (σ + b + 3) . σ −b−1

It can then be shown (see e.g. Sparrow, [Spa82]) that Re λ < 0 if r < r∗ , while Re λ > 0 if r > r∗ . It follows that if 1 < r < r∗ , the stationary points C± are both stable, and every orbit converges to one of these points. Thus, there is an attractor A, consisting of the points C− , C+ , and the unstable manifold Mu (O) connecting C− to C+ . More precisely, there is a value r1 ∈ ]1, r∗ [ such that: 1. If 1 < r < r1 , all three eigenvalues of J(C± ) are real negative; 2. If r1 < r < r∗ , there are two complex conjugate eigenvalues with negative real part. In this range, Mu (O) circles around C− and C+ ; 3. If r > r∗ , the two complex conjugate eigenvalues have positive real part, so the stationary points O, C+ and C− are all unstable (fig. 1.15).

z

20 y −20

20

x

Figure 1.15: Behavior of the orbits near the equilibrium points (unstable case).

36

1

Dynamical Processes

The numerical evidence confirms the existence of an attractor. Near C± , orbits arrive along the stable manifolds Ms (C± ) (corresponding to the real negative eigenvalue of J(C± )), and spiral out along the two-dimensional surface Mu (C± ). This behavior was first observed by Lorenz, in whose original computations the parameter values are σ = 10, b = 8/3; corresponding to these values, r∗ = 470/19 ≈ 24.74, and r1 ≈ 24.06. Lorenz’ so-called BUTTERFLY ATTRACTOR is observed at r = 28 (fig. 1.16).

z 20

10

y 10

−20

−10

−10 −20

10

x

Figure 1.16: The “butterfly” attractor.

1.6 Duffing’s Equation 1.6.1 The General Model The second example we consider is that of the so-called D UFFING EQUATION, which describes the motion of a vibrating spring subject to a nonlinear restoring term. The corresponding ODE model is determined in accord to Hooke’s law. Taking into account the dissipation effects due to friction (as measured by a numerical coefficient k), and assuming the presence of a periodic forcing term, the evolution of this system is governed by the nonlinear second order ODE x¨ + kx˙ + x3 − x = λ cos ω t ,

(1.52)

1.6

37

Duffing’s Equation

where k, λ and ω > 0. Equation (1.52) is equivalent to the first order system ( x˙ = y (1.53) y˙ = −ky + x − x3 + λ cos ω t , which is of type (1.17), with u = (x, y) ∈ X = R2 . It is not difficult to verify that (1.53) has, for each λ ∈ R and u0 ∈ R2 , a unique global solution u(·, u0 , λ ) ∈ C1 ([0, +∞[; R2 ). As was the case for the logistic equation, the asymptotic behavior of the solutions of system (1.53) is sharply influenced by the values of the parameter λ . When λ = 0, system (1.53) is autonomous, and the asymptotic behavior of its solutions can be studied with elementary techniques. In this case, (1.53) becomes ( x˙ = y (1.54) y˙ = x − x3 − ky . The stationary points of this system are the origin O, and the points C± := (±1, 0). To study the stability of these stationary points, we refer again to theorem A.7, and consider the characteristic polynomial of the linearized system, i.e. P(x, y; µ) := det[J(x, y) − µI] = µ(µ + k) + 3x2 − 1 . At the origin, P(0, 0; µ) = µ 2 + kµ − 1, which has two real distinct roots with op2 posite sign; √ thus, (0, 0) is a saddle point. At C± , P(±1, 0; µ) = µ + kµ + 2. Thus, if k > 2 2, there are two √ real, distinct, negative eigenvalues, and (±1, 0) are stable nodes. If instead k < 2 2, there are two complex conjugate eigenvalues. Since these have negative real part, both C+ and C− are stable sinks (fig. 1.17). The unstable y

C+ O x C−

Figure 1.17: The origin is an unstable saddle point; C+ and C− are stable sinks.

manifolds of O converge to the fixed points C± ; correspondingly, solutions of (1.54) with initial values on these manifolds converge to the stationary solutions (±1, 0).

38

1

Dynamical Processes

When λ 6= 0, the analogous of (1.54), i.e. system (1.53), is not autonomous, and will not have stationary solutions. However, since the system has a periodic forcing term, it may have periodic solutions, at least for some values of λ . In this case, we can carry out an analysis of the asymptotic behavior of the system by means of Poincaré maps. Namely, for each integer m we consider the sequence (un )n∈N , with un := u(2mnπω −1 , u0 , λ ). Each choice of m defines a stroboscopic map, and we can try to find the fixed points (if any) of these maps, and determine their stability. These fixed points would then correspond to periodic solutions of the system. For example, figure 1.18 refers to the stroboscopic map defined by system (1.53), with

1

−1

1

Figure 1.18: Attractor for Duffing’s equation.

λ = 0.5, k = 0.3 and ω = 1, corresponding to the case m = 1. Recall that this map is obtained by recording the projections of the points (u(2nπ, u0 , 0.5))n∈N . This stroboscopic map shows evidence of the existence of an attractor. We mention explicitly that the particular shape of this attractor depends on the choice of the sampling points tn ; for example, the sequence (u(2π(n + 1)ω −1 , u0 , 0.5))n∈N defines a different stroboscopic map, and produces an attractor with a different geometrical shape. More generally, numerical experiments show that for small values of λ there are at first periodic solutions with successive doublings of the period; then, at λ = 21 , the evolution of the system appears to be chaotic. We refer to Jordan-Smith, [JS87], and Guckenheimer-Holmes, [GH83], for a thorough examination of these cases.

1.6.2 A Linearized Model To better illustrate the procedure we have outlined for the nonautonomous case, we consider a simpler, linearized version of Duffing’s equation (1.52). More precisely, we consider the equation

x¨ + x˙ − 2x = −5(sint − cost) ,

1.6

39

Duffing’s Equation

which we transform as usual into the system ( x˙ = y y˙ = 2x − y − 5(sint − cost) .

(1.55)

As proposed, we consider the Poincaré sequence on the phase space X = R2 xn = x(2nπ) , yn = y(2nπ) ,

(1.56)

and look for fixed points of this sequence. The initial value problem for (1.55) with initial conditions x(0) = x0 , y(0) = y0 has the explicit solution ( x = 13 (2x0 + y0 )et + 31 (x0 − y0 + 3)e−2t + 2 sint − cost , (1.57) y = 13 (2x0 + y0 )et − 32 (x0 − y0 + 3)e−2t + 2 cost + sint (= x) ˙ . Correspondingly, the sequence (1.56) becomes ( xn = 13 (2x0 + y0 )e2nπ + 31 (x0 − y0 + 3)e−4nπ − 1 , yn = 13 (2x0 + y0 )e2nπ − 32 (x0 − y0 + 3)e−4nπ + 2 .

(1.58)

Let P = (−1, 2). We immediately see that P is a fixed point for the sequence (1.58) if and only if the initial conditions (x0 , y0 ) are such that 2x0 + y0 = 0 , x0 − y0 + 3 = 0 , i.e. if and only if (x0 , y0 ) = P. Corresponding to this choice of initial values, we see from (1.57) that (1.55) has the periodic solution x = 2 sint − cost , y = 2 cost + sint . To determine the stability of this solution, we study the stability of P as fixed point of the sequence (1.58). Setting Ms (P) = {(x, y) : 2x + y = 0}, Mu (P) = {(x, y) : x − y + 3 = 0} (these are two straight lines, intersecting at P), we easily verify that Ms (P) and Mu (P) are the stable and unstable manifolds of P (see definition 2.22 of section 2.3.3 below). Indeed, from (1.58) we have that (x0 , y0 ) ∈ Ms (P)

=⇒

(xn , yn ) → (−1, 2) as n → +∞ ,

while (x0 , y0 ) ∈ Mu (P)

=⇒ (xn , yn ) → (−1, 2) s M (P) ∪ Mu (P), then (x0 , y0 )

Equivalently, if (x0 , y0 ) ∈ (x(t), y(t))t ∈R of (1.57), with

as n → −∞ . belongs to a motion t 7→

(x(t), y(t)) → (−1, 2) Ms (P),

as t → +∞ if (x0 , y0 ) ∈ or as t → −∞ if (x0 , y0 ) ∈ Mu (P). If instead s u (x0 , y0 ) ∈ / M (P) ∪ M (P), the sequence ((xn , yn ))n∈N will not have a limit point, regardless of how close (x0 , y0 ) is to (−1, 2). Thus, the fixed point (−1, 2) is unstable.

40

1

Dynamical Processes

1.7 Summary We conclude this chapter with a brief summary of the main ideas of this introduction. We consider a DYNAMICAL SYSTEM, defined by an abstract autonomous ODE d u = F(u) , dt

u(t) ∈ X ,

(1.59)

where X is a Banach space and F ∈ C(X ; X ). System (1.59) is FINITE DIMEN SIONAL if dim X < +∞. Since the dynamical system (1.59) is autonomous, if the corresponding Cauchy problem is well posed in the large the differential equation (1.59) defines a CONTINUOUS SEMIGROUP S, also called a SEMIFLOW. We are interested in determining how the asymptotic properties of this semiflow, i.e. its behavior as t → +∞, depend on the initial values we attach to (1.59). If the system contains some numerical parameters, we are also interested in how these may influence the asymptotic behavior of the system. In many situations, it is possible to describe the long-time behavior of a dynamical system by means of a bounded ATTRACTOR, to which all the orbits converge as t → +∞, independently of where they originate. This attractor may be compact, and also have finite dimension. We would like to have criteria that allow us to recognize the existence of such an attractor, even if its structure may in general be quite complex, and very little may be known of its geometric or differential properties (although for many types of important physical examples there may be reasonable estimates on its dimension). On the other hand, there are favorable examples where we can in fact determine that the attractor is contained into a finite dimensional, exponentially attracting manifold of X (the so-called INERTIAL MANIFOLD), or at least into a compact, finite dimensional EXPONENTIAL ATTRACTOR. In these cases, since orbits converge exponentially fast to these sets, after a relatively short transient time the dynamics of the system are essentially governed by a finite system of ODEs. The number of the equations in this system is in general determined by the dimension of these sets. We would therefore like to identify some criteria that allow us to deduce the existence of such sets and, possibly, meaningful estimates on their dimension.

Chapter 2 Attractors of Semiflows

In this chapter we introduce the definitions of the SEMIFLOW associated to a dissipative autonomous dynamical system in a Banach space, and of the attractor of this semiflow. We discuss some of the most relevant properties of semiflows and their attractors. As we shall see, the ideas and results presented in this chapter are a natural generalization to the infinite dimensional case of many well known notions of the qualitative theory of ODEs (where the dimension of the space is finite). Throughout this chapter, X is a Banach space, with norm k · k and induced distance d; however, most of the results we establish also hold in general complete metric spaces. In the sequel, we will adopt the following notation. If E ⊆ R and a ∈ E, we set E≥a := {x ∈ E : x ≥ a} . For example, we have already referred to the set R≥0 = [0, +∞[ . Analogous definitions hold for the sets E>a , E≤a and E
2.1 Distance and Semidistance Since attractors are sets to which orbits converge, we need to recall the following facts on the distance and semidistance of two subsets of a metric space X , with distance d. DEFINITION 2.1 Given two subsets A, B ⊆ X and x ∈ X , we set d(x, B) := inf d(x, b) ,

(2.1)

b∈B

∂ (A, B) := sup d(a, B) = sup inf d(a, b) , a∈A

a∈A b∈B

dist(A, B) := max{∂ (A, B), ∂ (B, A)} .

(2.2) (2.3)

41

42

2

Attractors of Semiflows

We remark that the map (A, B) 7→ ∂ (A, B) is a SEMIDISTANCE; that is, ∂ is not symmetric, and the equality ∂ (A, B) = 0 doesn’t necessarily imply that A = B (take e.g. A ⊂ B). Moreover, we can even have ∂ (A, B) = 0 with A ⊃ B. This is for example the case when B is open, and A = B. However, we have PROPOSITION 2.2 Let A and B be subsets of X , such that ∂ (A, B) = 0. Then A ⊆ B. In particular, if B is closed, A ⊆ B. PROOF By (2.2), the condition ∂ (A, B) = 0 implies that inf d(a, b) = 0

b∈B

for all a ∈ A. Thus, for each a ∈ A we can find a sequence (bn )n∈N ⊂ B, such that d(a, bn ) → 0. This means that bn → a, so a ∈ B. From the last part of proposition 2.2, we immediately deduce COROLLARY 2.3 The restriction of the map (A, B) 7→ dist(A, B) to the family of closed subsets of X is a metric. PROOF We only have to show that dist(A, B) = 0 implies A = B for closed A, B ⊆ X . Indeed, (2.3) implies that both ∂ (A, B) = 0 and ∂ (B, A) = 0. The conclusion then follows from proposition 2.2. The distance defined in corollary 2.3 is known as the H AUSDORFF closed sets in X .

DISTANCE

of

2.2 Discrete and Continuous Semiflows 2.2.1 Types of Semiflows We start by defining various types of flows and semiflows, as follows. DEFINITION 2.4 Let T be one of the sets R, R≥0 , N, or Z. A SEMIFLOW on X is a family S := (S(t))t ∈T of continuous (but not necessarily linear) maps in X , i.e.

2.2

Discrete and Continuous Semiflows

43

such that for all t ∈ T , S(t) ∈ C(X , X ) , which satisfies the so-called

SEMIGROUP

(2.4)

conditions

S(0) = I , S(t + t 0 ) = S(t)S(t 0 ) ,

(2.5) (2.6)

for all t, t 0 ∈ T , and the additional continuity condition S(·)x ∈ C(T ; X )

(2.7)

for all x ∈ X . Furthermore: 1. If T is either R or Z, the semiflow is called a FLOW. 2. If T is either R or R≥0 , the flow (respectively, the semiflow) is called CONTIN UOUS . 3. If T is either Z or N, the flow (respectively, the semiflow) is called

DISCRETE .

Conditions (2.5) and (2.6) were already introduced in (1.20) and (1.21), and refer to the semigroup or group properties of S. S becomes a semiflow if it satisfies the additional requirements of the continuity of the maps x 7→ S(t)x

and

t 7→ S(t)x ,

respectively for all fixed t ∈ T and all fixed x ∈ X , as required in conditions (2.4) and (2.7). Note that the continuity in t, that is (2.7), is trivially satisfied for discrete flows and semiflows. Also, if S is a flow, S is in particular a GROUP of continuous operators on X . Alternatively, S may be called: 1. A CONTINUOUS SEMI - DYNAMICAL SYSTEM if T = R≥0 ; 2. A CONTINUOUS DYNAMICAL SYSTEM if T = R; 3. A DISCRETE SEMI - DYNAMICAL SYSTEM if T = N; 4. A DISCRETE DYNAMICAL SYSTEM if T = Z. The term “continuous” therefore distinguishes these systems from “discrete” ones, where the “time” set T is discrete, such as those considered in section 1.4. Likewise, the prefix “semi-” refers to the fact that we only consider nonnegative values of the time variable (continuous or discrete). In particular, a discrete semiflow S can be identified with the continuous mapping S˜ : X → X , defined by S˜ := S(1). In fact, S(n) = S˜n , for all n ∈ T .

44

2

Attractors of Semiflows

If T is either R≥0 or N, in definition 2.4 we do not require that S(t) be invertible for any t ∈ T . However, if S(t) is invertible for each t ∈ T , we can extend S to the parameter set T− := {−t : t ∈ T } ,

(2.8)

and therefore to a flow, by setting S(−t) := S(t)−1 for each t ∈ T . This definition extends (2.6) in a natural way; more precisely, we have PROPOSITION 2.5 1. Let T be either Z or R, and S = (S(t))t ∈T be a flow on X . Then for all t ∈ T , S(t) is invertible, and S(t)−1 = S(−t). 2. Let T be either N or R≥0 , and S = (S(t))t ∈T be a semiflow on X , such that S(t) is invertible for each t ∈ T . Define T− as in (2.8), and set T∗ := T ∪ T− . For t ∈ T∗ , define ( ˜ := S(t)

S(t) S(−t)−1

if t ∈ T , if t ∈ T− .

˜ t ∈T∗ is a flow in X . Then the family S˜ := (S(t)) PROOF The first claim is immediate, since for all t ∈ T S(t)S(−t) = S(t − t) = S(0) = I . To prove that S˜ is a flow, it is sufficient to show the semigroup property (2.6). This is immediate if both t, t 0 ∈ T . Assume then that t ∈ T and t 0 ∈ T− . We first show that for all x ∈ X , ˜ 0 )S(t)x ˜ ˜ 0 + t)x . S(t = S(t

(2.9)

We must distinguish two cases, according to whether t 0 + t ∈ T or not. If t 0 + t ∈ T , (2.9) reads S(−t 0 )−1 S(t)x = S(t 0 + t)x ,

(2.10)

and this is established by applying the operator S(−t 0 ) to both sides of (2.10), recalling that this operator is injective. If instead t 0 + t ∈ T− , (2.9) reads S(−t 0 )−1 S(t)x = S(−t 0 − t)−1 x .

(2.11)

2.2

45

Discrete and Continuous Semiflows

Let y ∈ X be such that S(−t 0 − t)y = x. Then, since −t 0 ∈ T , S(t)S(−t 0 − t)y = S(t)S(−t − t 0 )y = S(−t 0 )y . This means that y = S(−t 0 )−1 S(t)x; thus, (2.11) holds. We now show that also ˜ S(t ˜ 0 )x = S(t ˜ + t 0 )x S(t)

(2.12)

for all x ∈ X . Again, if t 0 + t ∈ T , (2.12) reads S(t)S(−t 0 )−1 x = S(t + t 0 )x .

(2.13)

Let z ∈ X be such that S(−t 0 )z = x. Then, since t ∈ T , S(t + t 0 )x = S(t + t 0 )S(−t 0 )z = S(t)z , which means that (2.13) holds. If instead t + t 0 ∈ T− , (2.12) reads S(t)S(−t 0 )−1 x = S(−t − t 0 )−1 x ,

(2.14)

and this is established by applying the operator S(−t − t 0 ), which is injective, to both sides of (2.14). To conclude the proof of proposition 2.5, we must still consider the case when both t, t 0 ∈ T− ; that is, we must show that for all x ∈ X , S(−t)−1 S(−t 0 )−1 x = S(−t − t 0 )−1 x = S(−t 0 )−1 S(−t)−1 x .

(2.15)

To prove the first of these identities, let y := S(−t −t 0 )−1 x, and z := S(−t)−1 x. Then, since both −t and −t 0 ∈ T , x = S(−t)z = S(−t − t 0 )y = S(−t)S(−t 0 )y .

(2.16)

Since S(−t) is injective, (2.16) implies that z = S(−t 0 )y; thus, since S(−t 0 ) is also invertible, y = S(−t 0 )−1 z = S(−t 0 )−1 S(−t)−1 x . This means that the first of (2.15) holds. The second identity is proven in the same way, exchanging the role of t and t 0 . In particular, if S is a semiflow generated by an evolution equation, the operators S(−t), t ≥ 0, will be defined whenever the equation can be uniquely solved “in the past”, i.e. for t < 0 as well. In particular, this requires that for each t > 0 the operator S(t) be bijective (see proposition 2.13 below). In the sequel, we shall adopt the following CONVENTION 2.6 The underlying time-parameter set for a semiflow or flow S is always denoted by T , i.e., S = (S(t))t ∈T . Moreover, in order to avoid unnecessary complications in formulas and sentences, we agree that if S is a semiflow, then all

46

2

Attractors of Semiflows

time variables like t, τ, θ , s run only in this time set T , or in the set T− := {−t ∈ T } if they are negative, or in the set T∗ := T ∪ T− if they can be positive or negative. For example, if T = N and (Ωs )s∈T is a family of subsets of X , the intersection \

Ωs

s≥τ

has to be understood as an abbreviation of \

Ωs .

s,τ∈N s≥τ

Similarly, when we write s ≤ −t with t ≥ 0, we understand that the inequality −s ≥ t holds in the set T , where both t and −s are. Finally, we recall that if T = R≥0 , then T− = R≤0 and T∗ = R.

2.2.2 Example: Lorenz’ Equations As we have seen in chapter 1, if the Cauchy problem for an autonomous evolution equation is well posed on R≥0 , it generates a continuous semiflow S on X , defined by the identification of the function t 7→ u(t, u0 ) =: S(t)u0 as the solution to the Cauchy problem ( u˙ = F(u) u(0) = u0 . As an example, we verify that Lorenz’ equations (1.50) define a semiflow on R3 . PROPOSITION 2.7 For all values of σ , r and b (not necessarily positive), the system of Lorenz’ equations (1.50) defines a continuous semiflow in X = R3 . PROOF 1. System (1.50) has at least a local solution, determined by standard existence and uniqueness results (see e.g. theorem A.1). That is, for each u0 := (x0 , y0 , z0 ) ∈ R3 there exist T (u0 ) > 0 and a unique, maximally defined solution u(·, u0 ) : [0, T (u0 )[ → R3 of the Lorenz’ equations (1.50), with u(0, u0 ) = u0 . This defines S(t) at least for t ∈ [0, T (u0 )[, by S(t)u0 := u(t, u0 ) . 2. We now prove that each such local solution can be extended to a global one. We achieve this, by establishing an A PRIORI ESTIMATE on each local solution u(·, u0 ), which shows that if T (u0 ) < +∞, then each function t 7→ u(t, u0 ) is bounded in

2.2

Discrete and Continuous Semiflows

47

the interval [0, T (u0 )[. This would allow us to continue the solution beyond T (u0 ), contradicting the fact that T (u0 ) is finite. Thus, the a priori estimate yields that T (u0 ) = +∞ for all u0 ∈ R3 . Setting u(t, u0 ) = (x(t, u0 ), y(t, u0 ), z(t, u0 )) , dropping the arguments t and u0 , and using the differential equations (1.50), we obtain d 2 |u| = 2xx˙ + 2yy˙ + 2z˙z = −2σ x2 + 2(r + σ )xy − 2y2 − 2bz2 = 2u Mu> , (2.17) dt where > denotes transposition, | · | denotes the Euclidean norm in R3 , and 

−σ  M :=  12 (r + σ ) 0

1 2 (r + σ )

−1 0

 0  0 . −b

Let Λ be the largest eigenvalue of M, i.e. q 1 2 2 Λ := max 2 (σ + r) + (σ − 1) − (σ + 1) , −b . Expressing u with respect to the basis of the eigenvectors of M, which are orthogonal, and can therefore be chosen to be orthonormal, we obtain from (2.17) that d |u(t, u0 )|2 ≤ 2 Λ |u(t, u0 )|2 . dt From this we conclude that for all t ∈ [0, T (u0 )[, |u(t, u0 )| ≤ |u0 |eΛt ≤ |u0 | max{1, eΛ T (u0 ) } .

(2.18)

This estimate shows that u(·, u0 ) is bounded in [0, T (u0 )[, as claimed. As we have discussed, from this it follows that the operators S(t) are defined for all t ≥ 0. The semigroup properties of S now follow from the fact that system (1.50) is autonomous, and condition (2.7) follows from the fact that the function t 7→ u(t, u0 ) is continuously differentiable. We also mention that later on (in proposition 2.65) we shall show that if σ and b are positive, we can obtain a better estimate than (2.18), that is, a bound on |u(t, u0 )| independent of t. This estimate would clearly allow us to show global existence at once. 3. We proceed then to prove the global well-posedness and the continuity of each operator S(t). In fact, we show that for each t > 0, S(t) is locally Lipschitz continuous, in the sense that for all t > 0 and all bounded subsets G ⊂ R3 , there exists L > 0 such that for all u0 , u0 ∈ G, |S(t)u0 − S(t)u0 | ≤ L|u0 − u0 | ,

(2.19)

48

2

Attractors of Semiflows

with L a continuous, increasing function of t and σG := sup |g|. To show this, set g∈G

(x(t), y(t), z(t)) := S(t)u0 , (x(t), ¯ y(t), ¯ z¯(t)) := S(t)u¯0 . Then, the difference S(t)u0 − S(t)u¯0 =: (ξ (t), η(t), χ(t)) solves the system  ˙   ξ = ση −σξ η˙ = rξ − η − xz + x¯ ¯z   ˙ χ = −bχ + xy − x¯y¯ . As before, dropping the argument t and using Schwarz’ inequality, we obtain d |S(t)u0 − S(t)u¯0 |2 = 2ξ ξ˙ + 2η η˙ + 2χ χ˙ dt = 2 −σ ξ 2 + (r + σ − z)ξ η − η 2 + yξ χ − bχ 2 ≤ 2 −σ ξ 2 + (r + σ )ξ η − η 2 − bχ 2 − 2zξ η + 2yξ χ ≤ 2 (C + |S(t)u0 |) |S(t)u0 − S(t)u¯0 |2 , where C := max{|σ |, |r + σ |, |b|}. Thus, recalling (2.18), ( 2(C + |u0 |) · |S(t)u0 − S(t)u¯0 |2 d 2 |S(t)u0 − S(t)u¯0 | ≤ dt 2(C + |u0 |eΛt ) · |S(t)u0 − S(t)u¯0 |2

if Λ ≤ 0 , if Λ > 0 .

Thus, after integration we obtain that if Λ ≤ 0, |S(t)u0 − S(t)u¯0 | ≤ |u0 − u¯0 | ·e|(C+{z|u0 |)t}

:= ϕ1 (t, |u0 |)

while, if Λ > 0, |S(t)u0 − S(t)u¯0 | ≤ |u0 − u¯0 | exp Ct + |u0 |Λ −1 (eΛt − 1) . | {z } := ϕ2 (t, |u0 |) This estimate shows that (2.19) holds, with L := max{ϕ1 (t, σG ), ϕ2 (t, σG )} .

We remark that the solution operator associated to system (1.53), i.e. to Duffing’s equation, is not a semiflow, but a so-called PROCESS, or TWO - PARAMETER SEMI FLOW . This is because the semigroup properties fail, due to the fact that the system is not autonomous. Nevertheless, it is possible to establish analogous a priori estimates, which would allow us to show that for all values of k, λ and ω (not necessarily positive), the solutions to the Cauchy problem for Duffing’s equation, corresponding to

2.3

Invariant Sets

49

arbitrary initial data, are globally and uniquely defined, and depend continuously on the initial data. Moreover, Duffing’s equation defines a discrete semiflow in R2 , by means of the stroboscopic maps described in section 1.4.1. We conclude this section by remarking that our whole discussion on the asymptotic behavior of the solutions to Lorenz’ and Duffing’s equations obviously requires that these solution be globally defined. For Lorenz’ equations, we established this fact by resorting to a continuation argument, which in turn was based on the possibility of obtaining a global bound on any local solution of the equations (see (2.18)). This procedure is fairly general, and estimates of this type are usually known as A PRIORI estimates. In section 2.9.1 at the end of this chapter we present some well known methods of obtaining a priori estimates.

2.3 Invariant Sets In this section we introduce a number of sets, which are invariant with respect to a semiflow or flow S = (S(t))t ∈T , with T as in definition 2.4. Such sets include the orbits, the ω-limit sets, and the stable and unstable manifolds of the semiflow. Since the “time-set” T can be either continuous or discrete, the definitions and properties we present hold for both continuous and discrete semiflows or flows. In the sequel, we often adopt the following notation. If B ⊆ X and T ∈ T , we set BT :=

[

S(t)B

(2.20)

t ≥T

(recall our convention 2.6). DEFINITION 2.8 Let S = (S(t))t ∈T be a semiflow on a Banach space X . A subset Y ⊆ X is POSITIVELY INVARIANT (respectively, NEGATIVELY INVARIANT) for S if for all t ≥ 0, S(t)Y ⊆ Y (respectively, S(t)Y ⊇ Y) ; Y is INVARIANT if for all t ∈ T , S(t)Y = Y . We immediately have PROPOSITION 2.9 Let S be a flow, and Y ⊆ X . Then Y is negatively invariant if and only if S(t)Y ⊆ Y for all t ≤ 0. PROOF Assume first that Y is negatively invariant. Let y ∈ Y and t ≤ 0, and set θ = −t. Then θ ≥ 0, so y ∈ S(θ )Y. Thus, there is x ∈ Y such that y = S(θ )x.

50

2

Attractors of Semiflows

Since S(t)y = S(t)S(θ )x = S(t + θ )x = S(0)x = x, we deduce that S(t)y ∈ Y. Thus, S(t)Y ⊆ Y. Conversely, assume this holds, and let y ∈ Y, θ ≥ 0. Then, t = −θ ≤ 0, so x := S(t)y ∈ Y. Then, y = (S(t))−1 x = S(−t)x = S(θ )x ∈ S(θ )Y. This shows that Y is negatively invariant. We know from ODE theory that simple examples of invariant sets are stationary points and periodic orbits. If S is a flow, other examples are given by complete orbits, and the stable and unstable manifolds of a stationary point.

2.3.1 Orbits DEFINITION 2.10 Let S = (S(t))t ∈T be a semiflow on X , and x ∈ X . 1. The FORWARD ORBIT originating (or starting) at x is the set γ+ (x) :=

[ t ≥0

{S(t)x} =

[

S(t){x} .

t ≥0

2. A BACKWARD ORBIT ending at x is the image im(u) of a function u : T− → X such that u(0) = x, and for all t ≥ 0 and s ≤ −t, u(t + s) = S(t)u(s) .

(2.21)

If the backward orbit ending at x is unique, we denote it by γ− (x). 3. A COMPLETE ORBIT through x is the image im(u) of a function u : T∗ → X such that u(0) = x, and (2.21) holds for all t ≥ 0 and s ∈ T∗ . If the complete orbit through x is unique, we denote it by γ(x). REMARK 2.11 1. The forward orbit γ+ (x) is the image im(u) of the function u ∈ C(T+ , X ) defined by u(t) := S(t)x , t ∈ T+ , where T+ := T ∩ [0, +∞[. 2. Let β = im(u), with u : T− → X satisfying (2.21), be a backward orbit ending at x, and let y ∈ β . Then x belongs to the forward orbit γ+ (y), starting at y. Indeed, there is τ ≥ 0 with y = u(−τ). By (2.21) we have x = u(0) = u(τ − τ) = S(τ)u(−τ) = S(τ)y ; hence, x ∈ γ+ (y). Moreover, (2.21) implies the continuity of u, since S(·)u(s) is continuous for s ∈ T− . 3. An analogous statement holds for complete orbits. Now, we can rephrase the definition of invariance of a set by saying that a set Y is positively invariant if every forward orbit starting from a point in Y is contained in

2.3

Invariant Sets

51

Y, and that it is invariant if for every point y in Y there is a complete orbit through y which is contained in Y. When there is no risk of confusion, it is common to refer to forward orbits simply as orbits; we shall indeed often do so. Note that in definition 2.10, the uniqueness of the backward orbit, or of the backward part of a complete orbit, is not implied. The following example illustrates the possible lack of uniqueness of backward orbits. Example 2.12 Consider the semiflow on X = [0, 1] generated by the logistic map of section 1.4.4, with λ = 4, that is, by the function S : [0, 1] → [0, 1] defined by S(x) = 4x(1 − x) (whose range is all of [0, 1]). We know that the point s4 = 34 is a fixed point of S; hence, the singleton γ1 := { 43 } is, trivially, a complete orbit through s4 . On the other hand, we easily verify that the sequence γ2 := (xn )n∈Z defined recursively by p for n ∈ N xn := 43 , x−n−1 := 21 (1 − 1 − x−n ) is such that x0 = s4 and S(xn ) = xn+1 for each n ∈ Z. Hence, this sequence is another complete orbit through s4 . Obviously, γ2 6= γ1 ; however, as expected, the forward parts of these two orbits coincide. The following result shows that the bijectivity of a semiflow is a sufficient condition for the uniqueness of backward and complete orbits. PROPOSITION 2.13 Assume that S is a semiflow on X , such that S(t) is invertible for all t ∈ T . Then backward and complete orbits through any x ∈ X are unique, and γ− (x) =

[

S(t)−1 {x} , γ(x) = γ− (x) ∪ γ+ (x) .

(2.22)

t ≥0

PROOF 1. Suppose β1 and β2 are two backward orbits ending at x. They are then the images of two functions u1 , u2 ∈ C(T− ; X ); therefore, given z ∈ β1 , there is s ∈ T− such that z = u1 (s). Then, t := −s ∈ T , and from (2.21) S(t)z = S(t)u1 (s) = u1 (t + s) = u1 (0) = x , as well as S(t)u2 (s) = u2 (t + s) = u2 (0) = x . Since S(t) is invertible, z = u2 (s); hence, z ∈ β2 . This shows that β1 ⊆ β2 . Changing the role of β1 and β2 proves that β1 = β2 . Thus, there is a unique backward orbit γ− (x) ending at x. 2. Next, let w : T− → X be defined by w(s) := S(−s)−1 x .

52

2

Attractors of Semiflows

Then w ∈ C(T− , X ), and w(0) = x. Moreover, for all t ≥ 0 and s ≤ −t, since w(t +s) = S(−t − s)−1 x we have that x = S(−t − s)w(t + s) , and therefore S(t)x = S(−s)w(t + s) . But since also x = S(−s)w(s), we deduce that S(t)x = S(t)S(−s)w(s) = S(−s)S(t)w(s) as well. Thus, (2.21) holds for w, because of the invertibility of S(−s). It follows that the image of w is a backward orbit ending at x; since we have proven that backward orbits are unique, the first of (2.22) follows. 3. The uniqueness of complete orbits and the second of (2.22) are proven similarly.

2.3.2 Limit Sets As we know from the theory of ODEs, a natural way to study the limit behavior of orbits as t → ±∞ in the phase space X is to study the topological properties of limit sets. Essentially, these are sets constructed from the union of forward or backward orbits (when these are unique), starting either from single points (limit sets of a point), or from (bounded) subsets of X . To study these sets, for s ≥ 0 we define the set Ys as in (2.20), and A−s :=

[

S(t)−1 Y ,

t ≥s

where for each t ∈ T we have set S(t)−1 Y := {x ∈ X : S(t)x ∈ Y}

(2.23)

(that is, the preimage of Y under S(t)). Note that if S(t) is invertible for all t ∈ T , then, by the first of (2.22), [ A−s ⊆ γ− (y) . y∈Y

Note, also, that if s ≥ s0 , Y s ⊆ Y s0

and A−s ⊆ A−s0 .

DEFINITION 2.14 Let S be a semiflow on X , and Y ⊆ X . The set ω(Y) :=

\ s≥0

Ys =

\[ s≥0 t ≥s

S(t)Y

2.3

53

Invariant Sets

is called the ω - LIMIT SET of Y (again, recall convention 2.6). The set \

α(Y) :=

A−s =

\[

S(t)−1 Y

s≥0 t ≥s

s≥0

(recall (2.23)) is called the α - LIMIT SET of Y (fig. 2.1).

S(t1 )x S(t2 )x

D1

x p

q

D2

α-limit set for points in annulus D1 ω-limit set for points in (D1 ∪ D2 ) \ {q} Figure 2.1: α- and ω-sets.

Note that ω- and α-limit sets are closed. If Y = {u0 }, we simply write ω(u0 ) to denote the limit set of the point u0 , and similarly for α(u0 ). We note explicitly that if S is a discrete (semi)flow, then ω(Y) =

\ [

Sm (Y) ,

S := S(1) .

n≥0 m≥n

We now prove a result which we will often use in the sequel; namely, that each point of an ω-limit set can be approximated by points that are on forward orbits starting in their defining sets Y. PROPOSITION 2.15 Let S be a semiflow on X , let Y ⊂ X , and z ∈ X . Then z ∈ ω(Y) if and only if there exist sequences (ym )m∈N ⊂ Y and (tm )m∈N ⊆ T , with tm → +∞, such that z = lim S(tm )ym . m→∞

PROOF Assume first that z ∈ ω(Y). Then z ∈ Ys for all s ≥ 0; thus, there exist sequences (zsk )k∈N from each Ys , converging to z as k → +∞. This means that for

54

2

Attractors of Semiflows

all ε > 0 and all s ≥ 0 there exists Kε,s ∈ N such that for all k ≥ Kε,s , d(z, zsk ) < ε. Choosing s = m ∈ N>0 , ε = m1 and k = K 1 ,m =: km , we deduce that for all m ∈ N>0 m

there exists zm ∈ Ym such that d(zm , z) < m1 (take zm = zm km ). Then, z = lim zm . Now, for each m ∈ N>0 , [ zm ∈ Y m = S(t)Y t ≥m

(recall convention 2.6). Thus, there are tm ≥ m and ym ∈ Y such that zm = S(tm )ym . Consequently, S(tm )ym converges to z, and tm ≥ m → +∞. Conversely, assume there exist sequences (ym )m∈N ⊂ Y and (tm )m∈N ⊂ T , such that tm → +∞ and z = lim S(tm )ym . Then for all s ≥ 0 there exists m such that tm ≥ s. Therefore, S(tm )ym ∈ S(tm )Y ⊆ Ys , and z ∈ Ys . This is true for all s ≥ 0, so z∈

\

Ys = ω(Y) .

s≥0

This ends the proof of proposition 2.15. For future reference, we note explicitly the following “discrete” version of proposition 2.15. PROPOSITION 2.16 Let S be a discrete semiflow on X , let Y ⊂ X , and z ∈ X . Then z ∈ ω(Y) if and only if there exist sequences (ym )m∈N ⊂ Y and (nm )m∈N ⊆ T , with nm → +∞, such that z = lim Snm ym , m→∞

(2.24)

where, we recall, S := S(1). A result analogous to proposition 2.15 holds for α-limit sets; that is, points in an α-limit set can be approximated by points which have the property that the orbits starting from these points all intersect Y. More precisely, PROPOSITION 2.17 Let S be a semiflow on X , Y ⊂ X , and z ∈ X . Then z ∈ α(Y) if and only if there exist sequences (zn )n∈N ⊂ X and (tn )n∈N ⊆ T such that tn → +∞, z = limn→∞ zn and yn := S(tn )zn ∈ Y. In particular, if S is a flow then z ∈ α(Y) if and only if there exist sequences (yn )n∈N ⊂ Y and (tn )n∈N ⊆ T such that tn → +∞ and z = limn→∞ S(−tn )yn . PROOF Assume first that z ∈ α(Y). Then z ∈ A−s for all s ∈ T , with s ≥ 0. As in the proof of proposition 2.15, this implies that for all n ∈ N>0 there is zn ∈ A−n such

2.3

Invariant Sets

55

that d(zn , z) < 1n . Then, z = limn→∞ zn . Now, for each n ∈ N, [

zn ∈ A−n =

S(τ)−1 Y .

τ ≥n

Thus, yn := S(tn )zn ∈ Y for some tn ≥ n. Conversely, assume that z = limn→∞ zn and tn → +∞, with yn = S(tn )zn ∈ Y. Then for all s ≥ 0 there exists n such that tn ≥ s. Since zn ∈ S(tn )−1 {yn } ⊆ S(tn )−1 Y ⊆ A−s , it follows that z = limn→∞ zn ∈ A−s . This is true for all s ≥ 0, and therefore z∈

\

A−s = α(Y) .

s≥0

This ends the proof of proposition 2.17. Propositions 2.15 and 2.17 can be applied to the case when Y contains a single point: COROLLARY 2.18 Let S be a semiflow and x ∈ X . Then ω(x) =

\[

{S(t)x} = {z ∈ X : ∃ tn → +∞ with S(tn )x → z} .

s≥0 t ≥s

If S is a flow, then α(x) =

\[

{S(t)−1 x} = {z ∈ X : ∃ tn → +∞ with S(tn )−1 x → z} .

s≥0 t ≥s

Another consequence of propositions 2.15 and 2.17 is the following useful property of ω- and α-limit sets, whose proof is immediate: COROLLARY 2.19 If Y ⊂ Z ⊂ X , then ω(Y) ⊂ ω(Z) and α(Y) ⊂ α(Z).

2.3.3 Stability of Stationary Points 1. In analogy to the theory of ODEs (compare to definition A.6), we adopt the following “natural” terminology. DEFINITION 2.20 STATIONARY POINT

Let S = (S(t))t ∈T be a semiflow on X . A point x ∈ X is a for S if S(t)x = x

A stationary point x of S is said to be:

for all t ∈ T .

56

2

Attractors of Semiflows

1. S TABLE, if for any neighborhood U of x there is a neighborhood V ⊂ U of x such that any motion starting in V is defined and contained in U for all t ≥ 0; 2. U NSTABLE, if it is not stable; 3. A SYMPTOTICALLY STABLE, if x is stable and there is a neighborhood V of x such that any motion starting in V converges to x, i.e. if for all x0 ∈ V, lim S(t)x0 = x

t →+∞

(see fig. 2.2, and compare to definition 1.14). Finally, if a flow S in Rn is defined

x0

x0 V

V U

U

Figure 2.2: Stability and Asymptotic stability.

by an autonomous ODE u˙ = F(u) ,

(2.25)

a stationary point of S is also called an EQUILIBRIUM POINT. Thus, equilibrium points are solutions of the algebraic equation F(x) = 0. We know from classical stability theory for ODEs (see theorem A.7) that, if F : Rn → Rn is of class C1 , and if all eigenvalues of the matrix A = F 0 (x) have nonzero real part, then the stability of equilibrium points x of (2.25) is related to the sign of the real part of the eigenvalues of A. Indeed, in this case the Hartman-Grobman theorem (see theorem A.8) states that, in a neighborhood of x, the qualitative behavior of the flow generated by (2.25) is the same as that of the flow S = (etA )t ∈R generated by the linearized equation u˙ = Au . This motivates the following

2.3

57

Invariant Sets

DEFINITION 2.21 1. Let S be the flow in Rn generated by the ODE (2.25). An equilibrium point x is HYPERBOLIC if all the eigenvalues of the matrix A = F 0 (x) have nonzero real part. 2. More generally, let X be a Banach space, and F : X → X be of class C1 . A fixed point x of F is HYPERBOLIC if the spectrum of the linear operator A = F 0 (x) does not intersect the unit circle {|z| = 1} of C. To see how these two notions of hyperbolicity of fixed points are related, consider a matrix A and, for each fixed t ∈ R, the linear operator F = etA in X = Rn . Then, F 0 (x) = etA for all x ∈ X , and the number Λ is an eigenvalue of etA iff Λ = eλt , with λ an eigenvalue of A. Let λ = a + ib. Then, |Λ | = eat = e(Re λ )t . Consequently, |Λ | = 6 1 iff Re λ 6= 0. In particular, by theorem A.7, hyperbolic equilibrium points x of (2.25) are either unstable (if Re λ > 0 for at least one eigenvalue λ of F 0 (x)), or asymptotically stable (if Re λ < 0 for all eigenvalues λ of F 0 (x)). We refer e.g. to Hirsch-Smale, [HS93], for more details. 2. We now recall the definitions of stable and unstable manifolds of a stationary point. DEFINITION 2.22 Let S be a semiflow, and x ∈ X be a stationary point for S. The STABLE MANIFOLD and the UNSTABLE MANIFOLD of x are the sets Ms (x) and Mu (x) defined, respectively, by Ms (x) := z ∈ X : lim S(t)z = x (2.26) t →+∞

and Mu (x) := z ∈ X : S(t)−1 z is defined for all t ≥ 0, and lim S(t)−1 z = x . t →+∞

In other words, Ms (x) consists of the points of X which are origins of forward orbits converging to x as t → +∞, while Mu (x) consists of the points of X which are end points of backward orbits, converging to x as t → −∞. The sets Ms (x) and Mu (x) are nonempty, since x ∈ Ms (x) ∩ Mu (x) (since x is a fixed point, trivially S(t)−1 x = x for all t ≥ 0). Moreover, we have PROPOSITION 2.23 Let x be a stationary point of a flow S. Then Mu (x) = {z ∈ X :

lim S(t)z = x} .

t →−∞

(2.27)

58

2

Attractors of Semiflows

If x is stable, Mu (x) = {x}. PROOF The first claim is immediate. As for the second, assume the contrary. There is then z ∈ Mu (x), with z 6= x. Thus, ε := 21 kx − zk > 0. Let U = B(x, ε), and determine a neighborhood V = B(x, δ ) of x, in accord with definition 2.20, such that for all y ∈ B(x, δ ) and all t ≥ 0, S(t)y ∈ B(x, ε). Since x = lim S(θ )z = lim S(−t)z = lim S(t)−1 z , t →+∞

θ →−∞

t →+∞

there is t0 ≥ 0 such that S−1 (t)z ∈ B(x, δ ) for all t ≥ t0 . Let y := S(t0 )−1 z. Then, y ∈ B(x, δ ), so S(t0 )y ∈ B(x, ε). Since S(t0 )y = z, we reach a contradiction. The definitions of stability and instability of a stationary point, and of their stable and unstable manifolds, come together in the following definition. DEFINITION 2.24 Let γ be a complete orbit of a flow S on X . Let xs , xu be two distinct stationary points of S, respectively stable and unstable. γ is a HETEROCLINIC ORBIT joining xs to xu if γ ⊆ Mu (xu ) ∩ Ms (xs ) . This terminology is justified by (2.22), (2.26) and (2.27), which imply that for any x ∈ γ, lim S(t)x = xu ,

t →−∞

lim S(t)x = xs .

t →+∞

For example, figure 2.3 shows the two heteroclinic orbits γ+ and γ− joining the γ = γ+ ∪ γ− = Mu (O) y γ− ⊆ Mu (O) ∩ Ms (C− ) C+ γ+ O

γ+ ⊆ Mu (O) ∩ Ms (C+ )

x shaded area is Ms (C+ )

C−

γ−

unshaded area is Ms (C− ) separatrices

Figure 2.3: Heteroclinic Orbits for Duffing’s Equation γ+ is the heteroclinic orbit joining O to C+ , γ− is the heteroclinic orbit joining O to C− .

2.3

59

Invariant Sets

unstable stationary point O of the flow generated by Duffing’s equations (1.54) to its asymptotically stable stationary points C+ and C− .

2.3.4 Invariance of Orbits and ω -Limit Sets As in the finite-dimensional case (i.e., for systems of ODEs), the stable and unstable manifolds of a stationary point are examples of invariant sets. More precisely: PROPOSITION 2.25 Let S be a semiflow on X , and x a fixed point of S. The stable manifold Ms (x) is positively invariant. If S(t) is invertible for all t > 0, the unstable manifold Mu (x) is negatively invariant. PROOF 1. Assume first that z ∈ S(t)Ms (x) for some t ≥ 0 in the time set T . Then, z = S(t)y for some y ∈ Ms (x). Since lim S(θ )y = x ,

θ →+∞

it follows that also lim S(θ )z = lim S(θ )S(t)y = lim S(θ + t)y = x .

θ →+∞

θ →+∞

t →+∞

This means that z ∈ Ms (x), and therefore that S(t)Ms (x) ⊆ Ms (x). 2. To show the negative invariance of Mu (x), fix t > 0 and z ∈ Mu (x). Since S(t) is invertible, y := S(t)−1 z is well defined in X . We claim that y ∈ Mu (x). Indeed, recalling proposition 2.5, lim S(θ )−1 y = lim S(θ )−1 S(t)−1 z = lim S(θ + t)−1 z = x .

θ →+∞

θ →+∞

θ →+∞

z = S(t)y ∈ S(t)Mu (x).

Consequently, In accord with definition 2.8, this means that Mu (x) is negatively invariant, as claimed. Other examples of invariant sets are the complete orbits of a semiflow: PROPOSITION 2.26 Let S be a semiflow on X , and x ∈ X . Then any complete orbit γ through x is invariant. If in addition S is a flow, then for all x, y ∈ X , γ(x) ∩ γ(y) 6= ∅ ⇐⇒ γ(x) = γ(y) .

(2.28)

PROOF 1. Let γ be a complete orbit, and z ∈ γ. Then γ is the image of a function u : T∗ → X satisfying (2.21), and there exists s ∈ T∗ such that z = u(s). Hence, by (2.21), for all t ≥ 0 z = S(t)u(s − t) ∈ S(t)γ .

60

2

Attractors of Semiflows

This implies that γ ⊆ S(t)γ for t ≥ 0. Conversely, if z ∈ S(t)γ, then z = S(t)u(s) for some s ∈ T∗ . Then, again by (2.21), z = u(t + s) ∈ γ, and S(t)γ ⊆ γ. 2. To prove (2.28), fix x and y ∈ X . By the characterization of complete orbits given by proposition 2.13, and recalling that S(t)−1 = S(−t), γ(x) =

[ t ∈T

{S(t)x} , γ(y) =

[

{S(t)y} .

t ∈T

Thus, if γ(x) ∩ γ(y) is nonempty, there are z ∈ γ(x) ∩ γ(y), and tx , ty ∈ T such that z = S(tx )x = S(ty )y . Hence, x = S(ty − tx )y ∈ γ(y) , and, in the same way, y = S(tx − ty )x ∈ γ(x) . Let now a ∈ γ(x). Then, a = S(ta )x for some ta ∈ T , and a = S(ta + ty − tx )y ∈ γ(y) . This means that γ(x) ⊆ γ(y). We prove in the same way that γ(y) ⊆ γ(x); hence, γ(x) = γ(y). The converse of the statement is obvious. We remark that (2.28) generalizes the familiar fact that two distinct orbits cannot intersect, unless they coincide. For flows generated by ODEs, this is a consequence of general uniqueness theorems. Other sets that are invariant, at least under certain conditions on S, are the ω-limit sets. In fact, it is precisely this invariance property, together with an additional “regularity” assumption on S, that will allow us to construct an attractor for the semiflow. We start by showing the positive invariance of ω-limit sets. PROPOSITION 2.27 Let S be a semiflow on X , and B ⊆ X . Then for all t ≥ 0 in T , ω(S(t)B) = ω(B), and S(t)ω(B) ⊆ ω(B). PROOF 1. Fix t ≥ 0 in T , and let z ∈ ω(S(t)B). By proposition 2.15, there are sequences (tn )n∈N ⊆ T and (zn )n∈N ⊆ S(t)B, such that tn → +∞, and S(tn )zn → z. For each n ∈ N, let xn ∈ B be such that S(t)xn = zn . Then, S(tn +t)xn → z, so z ∈ ω(B). This shows that ω(S(t)B) ⊆ ω(B). 2. Conversely, if z ∈ ω(B) there are sequences (tn )n∈N ⊆ T and (xn )n∈N ⊆ B, such that tn → +∞ and S(tn )xn → z. Since θn := tn −t → +∞, θn ≥ 0 for all large enough n. Since S(t)xn ∈ S(t)B and S(θn )S(t)xn = S(tn )xn → z ,

2.3

Invariant Sets

61

by proposition 2.15 again we conclude that z ∈ ω(S(t)B). This shows that ω(B) ⊆ ω(S(t)B). 3. The proof of the positive invariance of ω(B) is similar. Given z ∈ ω(B), let (tn )n∈N and (xn )n∈N be as above. Since τn := tn + t → +∞, S(τn )xn = S(t)S(tn )xn → S(t)z ; therefore, S(t)z ∈ ω(x). The second statement of proposition 2.27 shows that ω(B) is positively invariant. We will soon present various sufficient conditions for ω(B) to be invariant; these conditions all require some degree of compactness of different subsets of ω(B). Before proceeding, though, we need to explain what we mean when we say that two sets attract each other, and to recall the definition of relatively compact and precompact sets. DEFINITION 2.28 Let S be a semiflow on X . Given two subsets A, B ⊂ X , we say that A ATTRACTS B if lim ∂ (S(t)B, A) = 0 ,

t →+∞

(2.29)

where ∂ is the semidistance in X , introduced in (2.2). DEFINITION 2.29 A set C ⊂ X is said to be RELATIVELY COMPACT if its closure C is compact. Also, C is said to be PRECOMPACT if it can be completed into a compact set. Thus, in a Banach space precompact sets are automatically compact. We can then present the most restrictive condition for the invariance of ω-limit sets: THEOREM 2.30 Let S be a semiflow on X , and B ⊆ X be such that ω(B) is compact, and attracts B. Then ω(B) is invariant under S. PROOF By proposition 2.27, we only need to show that ω(B) ⊆ S(t)ω(B) for all t > 0. Let x ∈ ω(B), and (tn )n∈N ⊆ T , (xn )n∈N ⊆ B, such that tn → +∞ and S(tn )xn → x. Since the sequence (xn )n∈N is attracted by ω(B), and tn − t → +∞, lim d(S(tm − t)xm , ω(B)) = 0 .

m→∞

Recalling (2.1), this means that, given any ε > 0, there is m0 ∈ N such that for all m ≥ m0 , there is zm ∈ ω(B) with the property that d(S(tm − t)xm , zm ) < ε .

(2.30)

62

2

Attractors of Semiflows

Since ω(B) is compact, the sequence (zm )m∈N contains a subsequence (zmk )k∈N , converging to some element z ∈ ω(B). Then, (2.30) implies that S(tmk − t)xmk → z as k → ∞. Since S(t) is continuous, S(tmk )xmk = S(t)S(tmk − t)xmk → S(t)z ; but since S(tmk )xmk → x, we conclude that x = S(t)z ∈ S(t)ω(B). This shows that ω(B) ⊆ S(t)ω(B), as claimed. Theorem 2.30 can somewhat be relaxed, by assuming that ω(B) be only compact “asymptotically”: THEOREM 2.31 Let S be a semiflow on X . Assume there are a nonempty subset B ⊆ X and a number τ ≥ 0 such that the set Bτ (recall definition (2.20)) is relatively compact in X . Then the set ω(B) is nonempty, compact and invariant. If in addition B is connected by arcs and T is either R or R≥0 , ω(B) is connected by arcs. Moreover, for all bounded set G ⊆ B, lim ∂ (S(t)G, ω(B)) = 0 .

t →+∞

(2.31)

PROOF 1. Since B is nonempty, it contains an element x. The sequence (yn )n∈N defined by yn = S(n + τ)x takes values in Bτ , thus in the compact set Bτ . Hence, there is a subsequence (ynk )k∈N converging to a point y ∈ Bτ . By proposition 2.15, y ∈ ω(B), so ω(B) is not empty. 2. We know from proposition 2.27 that ω(B) is positively invariant; thus, to prove its invariance it is sufficient to show that ω(B) ⊆ S(t)ω(B) for all t > 0. Fix t > 0 and z ∈ ω(B), and let (zm )m∈N ⊂ B, (tm )m∈N ⊆ T be such that S(tm )zm → z, as per proposition 2.15. Let θm := tm − t. Then θm ≥ τ for all sufficiently large m, and S(θm )zm ∈ Bτ . Thus, there is a subsequence (S(θmk )zmk )k∈N converging to an element y ∈ Bτ . By proposition 2.15, y ∈ ω(B); since S(t) is continuous, S(t)y = lim S(tmk )zmk = z . k→+∞

Thus, z ∈ S(t)ω(B). This concludes the proof of the invariance of ω(B). 3. The compactness of ω(B) follows from the inclusions ω(B) =

\ s≥0

Bs ⊆

\

Bs ⊆ Bτ ,

s≥τ

the last set being compact. 4. If B is connected by arcs and if T is either R or R≥0 , then for all s ≥ 0 the set Bs is connected by arcs. Indeed, if u, v ∈ Bs , there exist tu ,tv ≥ s and x, y ∈ B such that u = S(tu )x and v = S(tv )x. Without loss of generality let tu ≤ tv . By assumption there is a continuous mapping p : [0, 1] → B with p(0) = x and p(1) = y. The mapping

2.3

63

Invariant Sets

τ 7→ S(tu + τ(tv − tu )p(τ), τ ∈ [0, 1], is continuous with values in Bs . Thus u and v are connected in Bs by the continuous arc {z ∈ X : z = S(tu + τ(tv − tu ))p(τ) , τ ∈ [0, 1]} . It follows that Bs , and therefore ω(B), is also connected. 5. Finally, to prove (2.31) we argue by contradiction. Thus, we assume there are: a bounded set G ⊆ B, a number ε0 > 0, and sequences (tn )n∈N ⊆ T , (gn )n∈N ⊆ G, such that tn → +∞, and for each n ∈ N, d(S(tn )gn , ω(B)) ≥ ε0 .

(2.32)

Since tn → +∞, there exists a number N such that tn ≥ τ for all n ≥ N. Thus, S(tn )gn ∈ S(tn )G ⊆ S(tn )B ⊆ Bτ . Since Bτ is relatively compact, there is a subsequence (S(tnk )gnk )k∈N , converging to a limit z ∈ Bτ . By proposition 2.15, z ∈ ω(B). From (2.32) we obtain then the contradiction 0 < ε0 ≤ d(S(tnk )gnk , ω(B)) = inf d(S(tnk )gnk , y) ≤ d(S(tnk )gnk , z) → 0 . y∈ω(B)

This concludes the proof of theorem 2.31. The importance of theorem 2.31 resides in the fact that it describes a class of sets, namely the sets ω(B), B ⊆ X , which share some of the properties we have stated to be characteristic of an attractor. Each of these sets is in fact nonempty, compact, invariant, and attracts some orbits (at least, the orbits starting from bounded subsets of B). Before proceeding, we mention that a similar result holds for α-limit sets. In fact, using the characterization of these sets given by proposition 2.17, we can prove THEOREM 2.32 Let S be a flow on X , and assume there is a nonempty subset Y ⊆ X such that S(−t)Y = 6 ∅ for all t ≥ 0. Assume further that there is a number τ > 0 such that the set [ S(−t)Y t ≥τ

is relatively compact in X . Then the set α(Y) is nonempty, compact and invariant, and for all bounded subset G ⊆ B, lim ∂ (S(t)G, α(Y)) = 0 .

t →−∞

We omit the proof of this result.

64

2

Attractors of Semiflows

2.4 Attractors In this section we present the definition and the main properties of the global attractor of a semiflow S on X . We only consider attractors that are compact, although it is possible to consider noncompact attractors.

2.4.1 Attracting Sets We start by recalling, from definition 2.28, that a subset A ⊆ X attracts another subset B ⊂ X if (2.29) holds, i.e. if lim ∂ (S(t)B, A) = 0 .

t →+∞

(2.33)

For example, (2.31) shows that if the semiflow S is uniformly compact for large t, the ω-limit set of a bounded set B attracts all orbits originating in B. Another, fundamental example of attracting set is given by THEOREM 2.33 Let S be a semiflow on X . Assume there is a nonempty subset B ⊆ X , which is compact and positively invariant. Then, the ω-limit set A := ω(B) is a compact, invariant set, which attracts all bounded sets G ⊆ B (including B itself). PROOF Since B is positively invariant and compact, for any τ ≥ 0 Bτ =

[ t ≥τ

S(t)B ⊆

[

B ⊆ B = B.

t ≥τ

Hence, Bτ is relatively compact in X , and we conclude by applying theorem 2.31. We proceed now to define the attractor of a semiflow: DEFINITION 2.34 A subset A ⊆ X is an ATTRACTOR for the semiflow S on X if A is compact and invariant, and there is a neighborhood U of A in X such that (2.33) holds for any bounded subset B ⊂ U. The largest neighborhood U of A such that (2.33) holds is called the BASIN OF ATTRACTION of A (fig. 2.4). In other words, A attracts all orbits starting sufficiently close to it. We remark that (2.33) carries no information on the rate of convergence of the orbits to the attractor. However, this rate is uniform for all orbits starting in the same set B; i.e., the rate of convergence does not depend on the particular initial value u0 , as long as u0 ∈ B. In chapter 4 we shall instead introduce another class of attracting sets, namely the EXPONENTIAL ATTRACTORS. As their name implies, the rate of

2.4

65

Attractors

γ

p

Figure 2.4: Basins of Attraction. convergence of orbits to these sets is exponential; the same will be true for the other class of attracting sets we introduce in chapter 5, that is, the INERTIAL MANIFOLDS. When X is infinite dimensional, we may have to work in subspaces X1 ,→ X with a stronger topology (one possible reason being, as we have remarked above, to show precompactness of one of the sets Bτ in X ). In this case, we give DEFINITION 2.35 Let X1 be a subspace of X , endowed with a stronger topology generated by a distance d1 in X1 . Define a corresponding semidistance ∂ 1 as in (2.2). A set A ⊆ X1 is an ATTRACTOR IN X1 for the semiflow S on X if A is an attractor of S in X , if it is invariant in X1 , and if there is a neighborhood U of A in X1 such that for any subset B ⊂ U, bounded in X1 , lim ∂ 1 (S(t)B, A) = 0 .

t →+∞

When a semiflow is defined by an evolution equation in X , verification of the invariance of A in X1 requires that orbits originating in X1 remain in X1 for all t > 0. For systems defined by PDEEs, this translates into a regularity property of the differential equation (see the examples of chapter 3).

2.4.2 Global Attractors In many situations, it is possible to show that a semiflow admits an attractor, to which all orbits converge. This motivates the following DEFINITION 2.36 Let S be a semiflow on X , and A ⊂ X an attractor for S. A is called a MAXIMAL (or GLOBAL, or UNIVERSAL) ATTRACTOR if its basin of attraction is all of X . We now give some preliminary results that characterize global attractors, and give some sort of justification to this terminology.

66

2

Attractors of Semiflows

PROPOSITION 2.37 Global attractors are unique. PROOF Assume that A1 and A2 are global attractors. Then they are both bounded and invariant, and attract all bounded subsets of X . In particular, they attract each other; thus, from (2.33), lim ∂ (S(t)A1 , A2 ) = lim ∂ (S(t)A2 , A1 ) = 0 .

t →+∞

t →+∞

The invariance of A1 and A2 implies then that ∂ (A1 , A2 ) = ∂ (A2 , A1 ) = 0. By (2.3) of definition 2.1, then, dist(A1 , A2 ) = 0. Since both A1 and A2 are compact, by corollary 2.3 it follows that A1 = A2 . Next, we show that global attractors are indeed maximal, with respect to set inclusion, among compact invariant sets of X (thus, in particular, among attractors). PROPOSITION 2.38 Let S be a semiflow on X . Let A2 be a compact invariant set, and assume that A1 ⊆ A2 is a global attractor. Then A1 = A2 . PROOF Since A1 ⊆ A2 , ∂ (A1 , A2 ) = 0. Since A2 is invariant, 0 ≤ ∂ (A2 , A1 ) = ∂ (S(t)A2 , A1 ) . Since A2 is bounded, it is attracted by A1 , so lim ∂ (S(t)A2 , A1 ) = 0 .

t →+∞

This implies that ∂ (A2 , A1 ) = 0. Thus, dist(A1 , A2 ) = 0; by corollary 2.3, it follows that A1 = A2 . Finally, we give a characterization of the structure of a global attractor, which will be of fundamental importance in chapter 3. PROPOSITION 2.39 Let S be a semiflow S on X , and assume that S admits a global attractor A. Let x ∈ X . Then x ∈ A if and only if there exists a complete orbit γ(x) through x, contained in A. PROOF Assume first that x ∈ A. Since A is invariant, the forward orbit γ+ := {S(t)x : t ≥ 0} is contained in A. To extend this orbit to t ≤ 0, we first remark that, since S(1)A = A, there exists y1 ∈ A such that S(1)y1 = x. Likewise, there exists

2.4

67

Attractors

y2 ∈ A such that S(1)y2 = y1 , and we can proceed inductively to construct a sequence (yn )n∈N ⊂ A, such that y0 = x , S(1)yn+1 = yn ,

n ≥ 0.

It is immediate to check that S(n)yn = x for all n ≥ 0. We now define a function u : T∗ → X by u(t) := S(t + n)yn , where n is any integer such that n ≥ −t. This function is well defined, because if also p ≥ −t, and for instance p > n, then yn = S(1)yn+1 = S(2)yn+2 = · · · = S(p − n)y p , and therefore S(t + n)yn = S(t + n)S(p − n)y p = S(t + p)y p . The function u is obviously continuous, and u(0) = x. By the invariance of A, the image γ := im(u) of u is included in A. If t ≥ 0, taking n = 0 we deduce that u(t) = S(t)y0 = S(t)x, which means that γ contains the forward orbit γ+ (x). To conclude that γ is in fact a complete orbit, we must verify the semigroup property (2.21), i.e. that u(t + s) = S(t)u(s) for all t ≥ 0 and s ∈ T∗ , s < −t. Given then t ≥ 0 and s < −t, choose an integer m ≥ −s. Then m ≥ −s − t as well, and u(s + t) = S(s + t + m)ym = S(t)S(s + m)ym = S(t)u(s) . Thus, γ is a complete orbit through x, contained in A. The converse statement of the theorem is obvious.

2.4.3 Compactness As theorem 2.33 shows, in order to use theorem 2.31 for the construction of an attractor it is desirable to start from a compact subset of X , which is positively invariant. Our next goal is to show that we can choose, as one such set, the closure of one of the sets Bτ introduced in theorem 2.31, for some τ ≥ 0. Thus, it is essential to verify that these sets are relatively compact, as required in theorem 2.31. To this end, if the system is finite dimensional it is sufficient to show that one such set Bτ is bounded in Rn . If instead the dimension of X is infinite, a sufficient condition is to show that there exists one such set Bτ which is bounded in a subspace X1 , compactly imbedded in X . This strategy is relatively common, and is the one we shall follow in chapter 3. With this in mind, we can easily determine a result, which can be immediately applied to show the required relative compactness of some set Bτ , in the case the semiflow is generated either by a finite dimensional system, or by an infinite dimensional system corresponding to evolution equations of “parabolic” type. DEFINITION 2.40 Let S be a semiflow on X . S is UNIFORMLY COMPACT FOR LARGE t if for any bounded set G ⊆ X there exists T > 0, depending on G, such that the set GT , defined as in (2.20), is relatively compact in X .

68

2

Attractors of Semiflows

The following result is then almost a restatement of theorem 2.31: COROLLARY 2.41 Assume the semiflow S on X is uniformly compact for large t. For any nonempty bounded set B ⊆ X , the set ω(B) is a nonempty, compact, invariant subset of X , such that for all bounded subset G ⊆ B, lim ∂ (S(t)G, ω(B)) = 0 .

t →+∞

2.5 Dissipativity Corollary 2.41 states that if the semiflow S is uniformly compact for large t, then ω-limit sets of bounded sets are attracting. Another key ingredient for the existence of a compact global attractor is the existence of an ABSORBING SET for the semiflow S. We shall in fact see that this set provides the natural “starting point” for the actual construction of the attractor of a semiflow. DEFINITION 2.42 Let S be a semiflow on X . A subset B ⊂ X is said to be ABSORBING , relative to a neighborhood U of B in X , if for all bounded sets G ⊂ U there exists T > 0, depending on G, such that S(t)G ⊆ B for all t ≥ T . In other words, all orbits originating in G enter B and, after possibly leaving it for a finite number of times, eventually remain in B forever. DEFINITION 2.43 Let S be a semiflow on X . S is called DISSIPATIVE, if it admits a nonempty, bounded absorbing set, relative to all of X . This definition of dissipativity coincides with the definition of bounded dissipativity (see e.g. Hale, [Hal88], where the notion of dissipativity is specialized for the attractors of points, bounded sets or compact sets). On the other hand, it may not agree with some other notions of dissipativity more common in physics. We now show that dissipativity is a natural property of semiflows that have a regularizing effect in time. PROPOSITION 2.44 Let S be a semiflow on X , uniformly compact for large t. Assume S has a nonempty, bounded absorbing set B ⊆ X , relative to all X . Then, S admits a compact, positively invariant absorbing set.

2.5

69

Dissipativity

PROOF Since B is bounded, by definition 2.40 there is T > 0 such that the set BT , as defined in (2.20), is relatively compact. We proceed then to show that the compact set BT is positively invariant and absorbing, relative to X . Let G ⊂ X be bounded. Since B is absorbing, there is T2 > 0 such that for all t ≥ T2 , S(t)G ⊆ B. Let T1 := T + T2 . For each t ≥ T1 , decompose t = θ + T2 . Then t − T2 = θ ≥ T , and therefore S(t)G = S(t − T2 )S(T2 )G ⊆ S(t − T2 )B ⊆ BT ⊆ B T . This shows that BT is absorbing. To show that BT is positively invariant, fix t ≥ 0 and x ∈ BT . There exists then a sequence (xk )k∈N ⊆ BT such xk → x. Since xk ∈ BT , for each k there is tk ≥ T such that xk ∈ S(tk )B; since t + tk ≥ tk ≥ T , S(t)xk ∈ S(t)S(tk )B = S(t + tk )B ⊆ BT . Then, by the continuity of the operator S(t), S(t)x = lim S(t)xk ∈ BT . Thus, BT is positively invariant. As we have stated, absorbing sets are a natural “starting point” for the construction of attractors. To justify this assertion, we show that the existence of an absorbing set is a necessary condition for the existence of a global attractor (recall that, in accord with definition 2.34, the attractors we consider are compact). PROPOSITION 2.45 Let S be a semiflow on X , admitting a global attractor A. Then S also admits a bounded, positively invariant absorbing set B, such that A ⊆ B. In particular, S is dissipative. PROOF Since A is compact, the set B1 :=

[

B(x, 1)

(2.34)

x∈A

is bounded, and therefore attracted by A. Thus, there is T1 > 0 such that for all t ≥ T1 , ∂ (S(t)B1 , A) = sup d(x, A) ≤ 12 . x∈S(t)B1

Since A is compact, for each t ≥ T1 and x ∈ S(t)B1 there is y ∈ A such that d(x, A) = d(x, y) ≤ 12 . Thus, x ∈ B(y, 1) ⊂ B1 , and we conclude that S(t)B1 ⊆ B1 if t ≥ T1 . Define then B2 :=

[

S(t)B1 , B := B1 ∪ B2 .

(2.35)

0≤t ≤T1

We claim that B is bounded, positively invariant, and absorbs all bounded sets of X . The boundedness of B follows from that of B1 and B2 , the latter being a consequence

70

2

Attractors of Semiflows

of the continuity of the map t 7→ S(t) on the compact interval [0, T1 ]. To show that B is positively invariant, fix t > 0 and z ∈ S(t)B. Then z = S(t)x for some x ∈ B. If x ∈ B2 , there are t2 ∈ [0, T1 ] and b ∈ B1 such that x = S(t2 )b; if x ∈ B1 , we set t2 = 0, b = x. In either case, z = S(t)x = S(t + t2 )b ∈ S(t + t2 )B1 ⊆ Bi ⊆ B , with i = 1 or i = 2 according to whether t + t2 ≥ T1 or not. Hence, S(t)B ⊆ B. To show that B is absorbing, consider any bounded set G ⊆ X . Repetition of the argument at the beginning of this proof shows that there is T > 0 such that for all t ≥ T , S(t)G ⊆ B1 ⊆ B. Finally, from (2.34) and the second of (2.35) we conclude that A ⊆ B1 ⊆ B. We conclude this section by pointing out that in definition 2.42, as well as in definitions 2.40 and 2.47 (below), the reference to bounded sets G can be interpreted in the spirit of our trying to control the effect of possible errors in the determination of the initial data, as mentioned in chapter 1. Indeed, even if the “true” initial value u0 of an orbit may only be known within a certain approximation, all these approximations will (hopefully!) lie in an explicitly identifiable bounded set of X . For example, when we approximate the number π up to three exact decimal digits, we are in fact considering numbers in the ball B(π, 10−3 ) of R (i.e., in the interval ]π − 10−3 , π + 10−3 [).

2.6 Absorbing Sets and Attractors We now come to the most important part of this chapter, where we show that if the semiflow S satisfies some compactness assumptions, the existence of an absorbing set is also sufficient for the existence of an attractor.

2.6.1 Attractors of Compact Semiflows Our first result concerns semiflows that have a regularizing effect in time. In the same spirit of corollary 2.41, we claim THEOREM 2.46 Assume that the semiflow S on X is uniformly compact for large t, and that it admits a nonempty, bounded absorbing set B, relative to a neighborhood U of B in X . Then the ω-limit set A := ω(B) is an attractor for S. More precisely, A is the global attractor for S in U. PROOF By the uniform compactness of S for large t, there is T0 > 0 such that the set BT0 , defined as in (2.20), is relatively compact in X . By corollary 2.41 we know

2.6

Absorbing Sets and Attractors

71

that A = ω(B) is nonempty, compact and invariant. We can now proceed along the lines of the proof of theorem 2.31 (we cannot apply this theorem directly, because we want to show that A attracts all the bounded subsets of U, and not just the subsets of B). Arguing by contradiction, assume that A does not attract all bounded sets of U. There is then a bounded set G ⊆ U for which ∂ (S(t)G, A) does not vanish as t → +∞. Thus, there also are δ > 0, and a sequence (tn )n∈N , such that tn → +∞ and for all n, ∂ (S(tn )G, A) = sup d(x, A) ≥ δ > 0 .

(2.36)

x∈S(tn )G

Since B is absorbing, there is T1 > 0 such that for all t ≥ T1 , S(t)G ⊆ B. Since tn → +∞, there also is n0 such that for all n ≥ n0 , tn ≥ T0 + T1 . By (2.36), for all n there is xn ∈ S(tn )G such that d(xn , A) ≥ δ2 . Let bn ∈ G be such that xn = S(tn )bn , and set zn := S(T1 )bn . Then zn ∈ S(T1 )G ⊆ B. If n ≥ n0 , then tn ≥ T1 and tn − T1 ≥ T0 ; therefore xn = S(tn )bn = S(tn − T1 )S(T1 )bn = S(tn − T1 )zn ∈ S(tn − T1 )B ⊆

[

S(t)B = BT0 .

t ≥T0

Thus, the sequence (xn )n∈N is precompact, and therefore admits a subsequence (xnk )k∈N , with xnk = S(tnk )bnk , converging to an element x¯ ∈ BT0 . Letting τk := tnk − T1 , we deduce as before that, as k → +∞, S(τk )znk = S(tnk − T1 )S(T1 )bnk = S(tnk )bnk = xnk → x¯ . Consequently, since znk ∈ B, x¯ ∈ ω(B) = A by proposition 2.15. On the other hand, d(x, ¯ A) = lim d(xnk , A) ≥ k→+∞

δ 2

,

and we reach a contradiction. It follows that A attracts all bounded subsets of U, as claimed. Finally, to show that A is the maximal attractor in U, assume that A1 ⊇ A is also a compact attractor of the bounded sets of U, as claimed. Since A1 is invariant and B is absorbing, as above we can determine T1 > 0, depending on A1 , such that for all t ≥ T1 , A1 = S(t)A1 ⊆ B. Thus, by corollary 2.19, ω(A1 ) ⊆ ω(B) = A . Since A1 is closed, its invariance also implies that ω(A1 ) =

\[ s≥0 t ≥s

S(t)A1 =

\[

A1 = A1 .

s≥0 t ≥s

We conclude then that A1 = ω(A1 ) ⊆ A and, therefore, A1 = A. In the next chapter we shall see how theorem 2.46 can be applied to establish the existence of attractors for the semiflow generated by a semilinear heat equation. In this case, this procedure is quite natural, since the requirement that the semiflow be uniformly compact for large t is a consequence of the smoothing effect of parabolic operators.

72

2

Attractors of Semiflows

2.6.2 A Generalization In contrast to the “parabolic” situation, the semiflow generated by semilinear dissipative wave equations does not have a smoothing effect for large t; therefore, theorem 2.46 cannot be directly applied to these equations. In this case, we can resort to a different type of result, in which the requirement that the semiflow be uniformly compact for large t can be relaxed, requiring that it be so only up to a uniformly decaying perturbation. More precisely, we give DEFINITION 2.47 Let S = (S(t))t ∈T be a family of continuous operators on X . S is UNIFORMLY DECAYING TO 0 if for all bounded sets G ⊆ X , lim sup d(S(t)x, 0) = 0 .

t →+∞ x∈G

(2.37)

We consider then a semiflow S = (S(t))t ∈T which admits a decomposition S = S1 + S2 ,

(2.38)

where S1 and S2 are families of continuous operators on X (i.e., they satisfy (2.7)), but not necessarily semiflows, and S1 , S2 are, respectively, uniformly compact for large t and uniformly decaying to 0. We first have PROPOSITION 2.48 Assume the semiflow S admits a decomposition as in (2.38), where the families of continuous operators S1 and S2 are, respectively, uniformly compact for large t and uniformly decaying to 0. Then for any sequences (xn )n∈N bounded in X , and (tn )n∈N ⊆ T , with tn → +∞, S(tn )xn converges ⇐⇒ S1 (tn )xn converges and S2 (tn )xn → 0 . In either case, lim S(tn )xn = lim S1 (tn )xn . PROOF Let (xn )n∈N and (tn )n∈N be as stated. Since (xn )n∈N is bounded, and S2 decays to 0, in accord to (2.37) lim sup d(S2 (t)xn , 0) = 0 .

t →+∞ n∈N

Thus, given ε > 0 there exists T > 0 such that for all t ≥ T and all n ∈ N, d(S2 (t)xn , 0) ≤ ε. Since tn → +∞, there is n0 such that tn ≥ T for all n ≥ n0 . Thus, for these n, d(S2 (tn )xn , 0) ≤ ε. This shows that S2 (tn )xn → 0. Then, S1 (tn )xn = S(tn )xn − S2 (tn )xn converges. The rest of the proof follows immediately. The next result generalizes part of corollary 2.41 to the case when S is not itself uniformly compact for large t, but is a uniformly vanishing perturbation of a uniformly compact family of operators.

2.7

Attractors via α-Contractions

73

PROPOSITION 2.49 Assume the semiflow S on X admits a decomposition as in proposition 2.48. Then for any nonempty bounded set B ⊆ X , the set ω(B) is nonempty, compact and invariant, and attracts B. PROOF We start by showing that ω1 (B) :=

\[

S1 (t)B = ω(B) .

s≥0 t ≥s

Indeed, if x ∈ ω(B), by proposition 2.15 there are sequences (xn )n∈N ⊂ B and (tn )n∈N ⊆ T , such that tn → +∞ and S(tn )xn → x. By proposition 2.48, S1 (tn )xn → x as well, so x ∈ ω1 (B). The converse is proven similarly, recalling that B is bounded. Now, since S1 is uniformly compact for large t, proceeding as in the proof of theorem 2.31 we deduce that ω(B) = ω1 (B) is nonempty and compact, and attracts B. The invariance of ω(B) is then a consequence of theorem 2.30. The importance of the next result is analogous to that of theorem 2.46, of which it can be seen as an extension. As we shall see in the next chapter, this theorem will in fact be applicable to certain types of dissipative evolution equations of hyperbolic type, to which theorem 2.46 cannot be applied, due to the lack of a regularizing effect of the corresponding solution operator. THEOREM 2.50 Assume that the semiflow S on X admits a decomposition (2.38) as in proposition 2.48, and that there is a nonempty, bounded absorbing set B, relative to a neighborhood U of B in X . Then the ω-limit set A := ω(B) is a compact attractor for S. More precisely, A is the global attractor for S in U. PROOF Since B is bounded, by proposition 2.49 its ω-limit set ω(B) = A is nonempty, compact and invariant. The rest of the proof proceeds exactly as in that of theorem 2.46, with the only difference that now only the sequence (S1 (tn )bn )n∈N can be said to be precompact. However, there still is a subsequence (S1 (tnk )bnk )k∈N , converging to some x¯ ∈ X , and proposition 2.48 guarantees that, in fact, S(tnk )bnk → x¯ as well.

2.7 Attractors via α -Contractions In this section we present an alternative procedure to show that the set ω(B) is a compact attractor, which does not require a decomposition like (2.38). Our presentation follows [EM93], where this method was used to show the existence of an

74

2

Attractors of Semiflows

attractor for the semiflow generated by a PDEE modelling the evolution of an extensible beam.

2.7.1 Measuring Noncompactness In the sequel, we denote by E a complete metric space, with distance d. If M ⊆ E, we denote by diam(M) the diameter of M, that is, diam(M) := sup{d(x, y) : x, y ∈ M} .

(2.39)

DEFINITION 2.51 Let A ⊆ E. A is TOTALLY BOUNDED if, given any ε > 0, it is possible to find a finite number of points {x1 , . . . , xN } in E such that A⊆

N [

B(xi , ε) .

i=1

This union is called a FINITE BALL - COVERING of A. Compactness can be expressed in terms of total boundedness, since a subset of E is compact if and only if it is complete and totally bounded. Thus, a closed subset can fail to be compact only if it is not totally bounded. Since we want to consider semiflows S which are not necessarily compact for large t, but do possess a bounded absorbing set, we need to somehow control the failure of compactness due to the lack of total boundedness. This can be done by means of the measures of compactness introduced by Kuratowski in [Kur66, ch. 3]. Given a subset M ⊆ E, we denote by I(M) the subset of R>0 consisting of all the positive numbers β such that M has a finite covering of sets, each having diameter not exceeding β . That is, β ∈ I(M) if and only if β > 0 and there are sets M1 , . . . , Mm ⊂ X , such that diam(M j ) ≤ β for j = 1, . . . , m, and M⊆

m [

Mj .

j=1

We can then define the notion of a measure of compactness: DEFINITION 2.52 Let 2E denote power set of E (that is, the set of its subsets). A MEASURE OF COMPACTNESS on E is the map α : 2E → [0, +∞] defined by ( +∞ if A has no finite covering , E ⊇ A 7→ α(A) := inf I(A) otherwise .

The main properties of measures of compactness are summarized in

2.7

Attractors via α-Contractions

75

PROPOSITION 2.53 Let α be a measure of compactness on E. Then 1. If A ⊂ E is bounded, α(A) < +∞; 2. If A ⊆ B, α(A) ≤ α(B) (monotonicity); 3. If α(A) = 0, then A is totally bounded; 4. If A1 ⊇ A2 ⊇ · · · ⊇ An ⊇ · · · is a decreasing sequence of nonempty closed sets such that α(An ) → 0 as n → +∞, then the set A :=

\

An

n≥1

is compact. PROOF 1) and 2) are immediate consequences of definition 2.52; note that I(A) ⊇ I(B) if A ⊆ B. To show 3), let ε > 0, and consider β ∈ I(A)∩ ]0, ε[. There is then a finite covering {C1 , . . . , Cn } of A, with diam(Ci ) ≤ β . Choosing points xi ∈ Ci , we see that the union of the balls B(xi , ε) covers A. Thus, A is totally bounded. Finally, to prove 4), note that, from 2), 0 ≤ α(A) ≤ α(An ) → 0 . Thus, α(A) = 0, so that, by 3), A is totally bounded. Since A is also complete, because it is closed, A is compact. We now introduce the notion of α-contraction, and the main result, which guarantees that ω(B) is indeed a compact attractor for the semiflow S. DEFINITION 2.54 Let B ⊆ E. A continuous map T : B → B is an α - CONTRAC TION on B, if there exists a number q ∈ ]0, 1[ such that for every subset A ⊆ B, α(T (A)) ≤ q α(A) .

(2.40)

We have then the following fundamental result: THEOREM 2.55 Assume that B ⊆ E is closed and bounded, and that T : B → B is an α-contraction on B. Consider the semiflow generated by the iterations of T , i.e. S = (T n )n∈N . Then ω(B), if nonempty, is a compact, invariant set, which attracts B. PROOF 1. For n ∈ N, set An := T n (B). Clearly, An ⊇ An+1 for each n. We show that, as a consequence, \ ω(B) = An =: A . n≥0

76

2

Attractors of Semiflows

To see this, note first that, since for all n ∈ N we obviously have T n (B) ⊆

[

T m (B) ,

m≥n

we immediately deduce that A ⊆ ω(B) =

\ [

T m (B) .

n≥0 m≥n

Conversely, let z ∈ ω(B). By proposition 2.16 (see (2.24)), there are sequences (n j ) j∈N and (z j ) j∈N ⊆ B, such that n j → ∞ and T n j z j → z as j → ∞. Now, for each n ∈ N there is jn ∈ N such that n j ≥ n for all j ≥ jn . Hence, for j ≥ jn , T n j z j ∈ T n j (B) ⊆ An j ⊆ An . Letting j → ∞, it follows that z ∈ An for all n ∈ N. Consequently, z ∈ A, and ω(B) = A. 2. Since B is bounded, by (1) of proposition 2.53 there is M > 0 such that α(B) ≤ M. A repeated application of (2.40) yields then α(An ) = α(T n (B)) ≤ qn α(B) ≤ qn M ; thus, α(An ) → 0. Since each An is closed, part (4) of proposition 2.53 implies that ω(B) = A is compact. 3. To see that ω(B) attracts B, we show that lim d(T n x, ω(B)) = 0 ,

n→∞

uniformly in x ∈ B; that is, we claim that for all ε > 0 there exists N such that for all integer n ≥ N and all x ∈ B, d(T n x, ω(B)) < ε . Proceeding by contradiction, assume there is ε0 > 0 such that for all integers j it is possible to find another integer n j ≥ j, and a point x j ∈ B, such that d(T n j (x j ), ω(B)) ≥ ε0 .

(2.41)

This process defines a bounded sequence ζ∗ := (T n j x j ) j∈N ⊂ B. If we can show that ζ∗ contains a convergent subsequence, we reach the desired contradiction, because by (2.41) the limit z of this subsequence would on the one hand be in ω(B) (by proposition 2.15), and on the other would satisfy d(z, ω(B)) ≥ ε0 . To show that ζ∗ does contain a convergent subsequence, let Σ be the subset of B consisting of all the sequences of the form ζ = (T m j x j ) j∈N , with x j ∈ B, m j ∈ N and m j → ∞ as j → ∞. Since α(ζ ) ≤ α(B) for all ζ ∈ Σ , 0 ≤ α0 := sup α(ζ ) < +∞ . ζ ∈Σ

2.7

Attractors via α-Contractions

77

We claim that α0 = 0. Otherwise, we could first choose a number θ such that 0 < θ < (1−q)α0 , and then a sequence ζ0 ∈ Σ such that α0 −θ < α(ζ0 ). Let ζ0 = (T m j x j ) j∈N . Since m j → ∞, there is j0 ∈ N such that m j ≥ 1 for all j ≥ j0 . Consider then the sequence ζ1 := (T m j −1 x j ) j≥ j0 . Since ζ1 can be written as ζ1 = (T nk yk )k∈N , with nk = m j0 +k − 1 → ∞ as k → ∞, and yk = x j0 +k ∈ B, it follows that ζ1 ∈ Σ ; therefore, α(ζ1 ) ≤ α0 . Next, setting ζ˜0 := T ζ1 = (T m j x j ) j≥ j0 , we see that the sequence T ζ1 coincides with the sequence ζ0 , deprived of its first j0 terms. We now check that dropping this finite number of terms does not affect the measure of α-compactness of ζ0 . Indeed, from part (2) of proposition 2.53 we first have α(ζ˜0 ) ≤ α(ζ0 ) . To show the opposite inequality it is sufficient to show that I(ζ˜0 ) ⊆ I(ζ0 ) . Now, if β ∈ I(ζ˜0 ) and C1 , . . . , Cr is a finite covering of ζ˜0 , such that diam(Ci ) ≤ β , the addition to this covering of the j0 balls B(T xi , 21 β ), 0 ≤ i ≤ j0 , produces a finite covering of ζ0 with sets whose diameter does not exceed β . Thus, β ∈ I(ζ0 ), as claimed. In conclusion, we have the chain of inequalities α0 − θ < α(ζ0 ) = α(ζ˜0 ) = α(T ζ1 ) ≤ q α(ζ1 ) ≤ qα0 < α0 − θ , which yields a contradiction. This means that α0 = 0 and, therefore, α(ζ ) = 0 for all ζ ∈ Σ . In particular, α(ζ∗ ) = 0, which implies, by part (3) of proposition 2.53, that ζ∗ is totally bounded. Hence, ζ∗ is compact, and contains a convergent subsequence, as claimed. Finally, the invariance of ω(B) follows from theorem 2.30 (in its discrete version). This concludes the proof of theorem 2.55. We now proceed to extend theorem 2.55 from the discrete case to the continuous one. The corresponding result will then provide the desired alternative to theorems 2.46 and 2.50. THEOREM 2.56 Assume that S is a continuous semiflow on X , admitting a bounded, positively invariant absorbing set B, and that there exists t∗ > 0 such that the operator S∗ := S(t∗ ) is an α-contraction on B. Let A∗ :=

\ [

S∗m (B) = ω∗ (B)

n≥0 m≥n

be the ω-limit set of B under the map S∗ , and set A :=

[ 0≤t ≤t∗

S(t)A∗ .

78

2

Attractors of Semiflows

Assume further that for all t ∈ [0,t∗ ], S(t) is Lipschitz continuous from B into B, with Lipschitz constant L(t), L : [0,t∗ ] → ]0, +∞[ being a bounded function. Then A = ω(B), and this set is the global attractor of S in B. PROOF 1. To show that A is compact, we note that the function F : [0, +∞[ ×B → B defined by F(t, x) := S(t)x is continuous on [0,t∗ ] × A∗ . To see this, we set L∗ := sup L(t) ,

(2.42)

0≤t ≤t∗

and fix (t0 , x0 ) ∈ [0,t∗ ] × A∗ . Since the map t 7→ S(t)x is continuous for each x ∈ X , given η > 0 there is δ1 > 0 such that if |t − t0 | ≤ δ1 , d(S(t)x0 , S(t0 )x0 ) ≤ 21 η ;

(2.43)

note that δ1 depends on η and, possibly, on (t0 , x0 ). Let δ := min{δ1 , 21 L∗−1 η} , then, if (t, x) is such that (d(x, x0 ))2 + |t − t0 |2 ≤ δ 2 , by (2.43) we have that d(S(t)x, S(t0 )x0 ) ≤ d(S(t)x, S(t)x0 ) + d(S(t)x0 , S(t0 )x0 ) ≤ L(t) d(x, x0 ) + d(S(t)x0 , S(t0 )x0 ) ≤ L∗ δ + 21 η ≤ η . This shows the continuity of F. It is then immediate to verify that A = A1 := F([0,t∗ ] × A∗ ) ; thus, A is compact, because F is continuous and [0,t∗ ] × A∗ is compact. 2. We show that A attracts all bounded subsets of B. Let G ⊆ B be bounded, and fix t ≥ t∗ . Given any x ∈ S(t)G and a∗ ∈ A∗ , let g ∈ G be such that x = S(t)g, and decompose t = nt∗ + θt , for suitable n ∈ N and θt ∈ [0,t∗ ]. Let a¯ := S(θt )a∗ . Then, a¯ ∈ A, and recalling (2.42) we can estimate d(x, a) ¯ = d(S(θt + t − θt )g, S(θt )a∗ ) ≤ L∗ d(S(t − θt )g, a∗ ) ≤ L∗ d(S(nt∗ )g, a∗ ) = L∗ d(S∗n g, a∗ ) . From this, it follows that inf d(x, a) ≤ d(x, a) ¯ ≤ L∗ d(S∗n g, a∗ )

a∈A

and, since a∗ is arbitrary in A∗ , inf d(x, a) ≤ L∗ inf d(S∗n g, a∗ ) .

a∈A

a∗ ∈A∗

(2.44)

2.7

79

Attractors via α-Contractions

Since g ∈ G ⊆ B, and B is positively invariant, S∗n g ∈ B. Thus, recalling the definition (2.2) of semidistance, we can proceed from (2.44) with inf d(x, a) ≤ L∗ sup

a∈A

inf d(b, a∗ ) = L∗ ∂ (S∗n B, A∗ ) .

n B a∗ ∈A∗ b∈S∗

(2.45)

Since (2.45) is true for arbitrary x ∈ S(t)G, it follows that sup inf d(x, a) = ∂ (S(t)G, A) ≤ L∗ ∂ (S∗n B, A∗ ) .

x∈S(t)G a∈A

(2.46)

Since A∗ attracts B under S∗ , (2.46) implies that A attracts G under S, as claimed. 3. We now show that A = ω(B). Let a ∈ A. There are then θ ∈ [0,t∗ ] and a∗ ∈ A∗ , such that a = S(θ )a∗ . Since A∗ = ω∗ (B), by proposition 2.16 there are sequences m (m j ) j∈N ⊆ N and (z j ) j∈N ⊆ B, such that m j → ∞ and S∗ j z j → a∗ as j → ∞. Let t j := θ + m j t∗ . Then, t j → ∞, and a = S(θ )a∗ = lim S(θ + m j t∗ )z j = lim S(t j )z j . j→∞

j→∞

Thus, again by proposition 2.15, a ∈ ω(B). This proves that A ⊆ ω(B). Conversely, let z ∈ ω(B). Then, there are sequences (t j ) j∈N ⊆ T and (z j ) j∈N ⊆ B, such that t j → ∞ and S(t j )z j → z as j → ∞. For each j ∈ N, we can write t j = m j t∗ + θ j , with m j ∈ N, θ j ∈ [0,t∗ ], and m j → ∞ as j → ∞. Since B is positively invariant, S(θ j )z j =: z˜ j ∈ B for all j. Hence, m

z = lim S(t j )z j = lim S(m j t∗ )S(θ j )z j = lim S∗ j z˜ j . j→∞

j→∞

j→∞

This means that z ∈ ω∗ (B) = A∗ . Since A∗ ⊆ A, it follows that ω(B) ⊆ A. Thus, A = ω(B). 4. Since A is compact and attracts B, and ω(B) = A, theorem 2.30 implies that A is invariant. This ends the proof of theorem 2.56. In conclusion, theorem 2.56 provides an alternative way to establish the existence of an attractor for a continuous semiflow, if we can choose t∗ so that the operator S(t∗ ) is an α-contraction, and S(t) is a Lipschitz continuous map on X , for all t ∈ [0,t∗ ]. In applications, this is often achieved by means of the intermediate results described next, which we will use for the dissipative evolution equations we consider in chapter 3.

2.7.2 A Route to α -Contractions In the light of our previous remark, it is clearly of interest to be able to give some sufficient conditions for a continuous map T : B → B, defined on a closed, bounded set B ⊂ X , to be an α-contraction. To this end, note first that this is obviously the case if T is a strict contraction, and also if T is compact, because in this case α(T (B)) = 0, and we can take arbitrary q ∈ ]0, 1[ in (2.40). This is exactly the situation we had in

80

2

Attractors of Semiflows

theorems 2.46 and 2.50. In fact, in both these cases we can choose T = S(t∗ ), with t∗ > 0 arbitrary in theorem 2.46 (where T is compact), and, in theorem 2.50, t∗ so large that S2 (t∗ ) is a strict contraction. We now show that a combination of these two effects still produces an α-contraction. DEFINITION 2.57 A pseudometric δ in X is PRECOMPACT with respect to the topology induced by the metric d of X if every sequence which is bounded relatively to the distance d has a Cauchy subsequence relative to δ . PROPOSITION 2.58 Let δ be a precompact pseudometric, and let B ⊂ X be bounded. Then B is totally bounded with respect to δ . That is, for all ε > 0 there is a finite covering of B, consisting of δ -balls Bδ (xi , ε) := {y ∈ X : δ (xi , y) < ε} ,

i = 1, . . . , m .

PROOF We proceed by contradiction. If there were ε0 > 0 such that there is no finite covering of B by means of δ -balls B1 = Bδ (x1 , ε0 ), . . . , Bm = Bδ (xm , ε0 ), arguing inductively we could construct a sequence (xi )i∈N , contained in B, and therefore bounded, such that xm+1 ∈ / Bm for each m. In particular, this would imply that δ (xm , xm+1 ) ≥ ε0 . Consequently, (xm )m∈N could not be a Cauchy sequence relative to δ . But this sequence must be a Cauchy sequence, since it is bounded and δ is precompact. We now show that if T fails to be contractive only because of a precompact pseudometric, it is still an α-contraction. PROPOSITION 2.59 Let B ⊂ X be bounded, δ a precompact pseudometric in X , and T : B → B be a continuous map. Suppose T satisfies the estimate d(T x, Ty) ≤ qd(x, y) + δ (x, y)

(2.47)

for all x, y ∈ B and some q ∈]0, 1[ independent of x and y. Then T is an α-contraction. PROOF We will show that (2.40) holds, with q from (2.47). Let A ⊆ B, and α0 = α(A) (note that α(A) ≤ α(B) < +∞, because B is bounded). By proposition 2.58, given ε > 0 we can cover B with a finite number of δ -balls B1 , . . . , Bn of radius ε. By the definition of α0 , we can also cover A with a finite number of sets C j , j = 1, . . . , m, with diam(C j ) ≤ α0 + ε for all j. Clearly, T (A) ⊆

[ i=1,...,n j=1,...,m

T (Bi ∩ C j ) .

2.8

Fractal Dimension

81

Recalling (2.47), we can estimate the diameter of each set of this covering of T (A) as follows. For x, y ∈ Bi ∩ C j we compute that d(T x, Ty) ≤ qd(x, y) + δ (x, y) ≤ q(α0 + ε) + ε ≤ qα0 + 2ε . This means that diam(T (Bi ∩ C j )) ≤ qα0 + 2ε = qα(A) + 2ε . Since ε is arbitrary, we conclude that α(T (A)) ≤ sup diam(T (Bi ∩ C j )) ≤ qα(A) , i, j

from which (2.40) follows.

2.8 Fractal Dimension As we have mentioned in chapter 1, in many situations it is possible to show that the attractor of a semiflow has finite dimension. Since in general attractors have a highly irregular structure (they may well be fractal sets), it is often necessary to consider their FRACTAL DIMENSION, whose definition we briefly recall in this section. Following [Fal85, Fal90, EFNT94], we introduce the following DEFINITION 2.60 Let X be a separable Hilbert space, and K ⊂ X be a compact subset. For δ > 0, denote by Nδ (K) the smallest number of sets of diameter at most equal to δ which can cover K. The FRACTAL DIMENSION of K is the number dimF (K) := lim sup δ →0

ln Nδ (K) . − ln δ

(2.48)

(We include the possibility that dimF (K) = +∞ for some sets K.) We recall that dimF (K), as defined in (2.48), is also known as the UPPER BOX dimension of K. There are corresponding definitions of lower boxcounting dimension and of box-counting dimensions of K, obtained by replacing, in (2.48), lim sup respectively by lim inf and lim. The fractal dimension of a set is always larger than or equal its Hausdorff dimension. Further properties of the fractal dimension are given in the following proposition. COUNTING

PROPOSITION 2.61 Let dimF be defined as in (2.48).

82

2

Attractors of Semiflows

1. If K1 , K2 are two compact sets of X , then: K1 ⊆ K2 =⇒ dimF (K1 ) ≤ dimF (K2 )

(2.49)

(i.e., the fractal dimension is monotonic). Moreover, dimF (K1 ∪ K2 ) ≤ max(dimF (K1 ), dimF (K2 )) , dimF (K1 × K2 ) ≤ dimF (K1 ) + dimF (K2 ) .

(2.50) (2.51)

2. If f : X → X is Lipschitz continuous, then for any compact set K ⊂ X , dimF ( f (K)) ≤ dimF (K) .

(2.52)

3. If M ⊆ RN is a smooth, N-dimensional compact manifold of RN , then dimF (M) = N . PROOF 1. The monotonicity of dimF is immediate, since if K1 ⊆ K2 , then any covering of K2 is automatically a covering of K1 . In particular, Nδ (K1 ) ≤ Nδ (K2 ), and (2.49) follows after division of this inequality by − ln δ , which is positive if 0 < δ < 1. 2. To prove (2.50), fix δ ∈ ]0, 1[ and cover K1 and K2 with exactly Nδ (K1 ) and Nδ (K2 ) balls of diameter δ . Then, the union of these coverings is a covering of K1 ∪ K2 , consisting of Nδ (K1 ) + Nδ (K1 ) balls of diameter δ . Since Nδ (K1 ∪ K2 ) is the minimum number of balls of diameter at most δ that are necessary to cover K1 ∪ K2 , it follows that, setting N˜ δ := max(Nδ (K1 ), Nδ (K2 )), Nδ (K1 ∪ K2 ) ≤ Nδ (K1 ) + Nδ (K2 ) ≤ 2N˜ δ . Thus, ln Nδ (K1 ∪ K2 ) ln 2 + ln N˜δ ≤ , − ln δ − ln δ from which (2.50) follows, letting δ → 0. 3. To prove (2.51), fix δ ∈ ]0, 12 [, and cover K1 and K2 with exactly Nδ (K1 ) and Nδ (K2 ) balls of diameter δ . Then, the product of these coverings is a covering of K1 × K2 , consisting of Nδ (K1 ) · Nδ (K1 ) balls of diameter at most 2δ . It follows that N2δ (K1 × K2 ) ≤ Nδ (K1 ) · Nδ (K2 ) , and, therefore, ln N2δ (K1 × K2 ) ln Nδ (K1 ) + ln Nδ (K2 ) ≤ . − ln(2δ ) − ln(2δ ) Multiplying and dividing the right side of this inequality by − ln δ , and letting δ → 0, we obtain (2.51).

2.9

A Priori Estimates

83

4. To prove (2.52), fix again δ ∈ ]0, 1[ and cover K with exactly Nδ (K) balls of diameter δ . Then f (K) can be covered by the images of these balls, which have diameter at most Lδ . Hence, NLδ ( f (K)) ≤ Nδ (K) and, therefore, dimF ( f (K)) = lim sup Lδ →0

= lim sup δ →0

ln Nδ (K) ln δ ln NLδ ( f (K)) ≤ lim sup − ln(Lδ ) ln L + ln δ − ln δ δ →0 ln Nδ (K) = dimF (K) . − ln δ

Thus, (2.52) holds.

2.9 A Priori Estimates The discussion of Lorenz’ equations in section 2.2.2 highlighted the importance of obtaining A PRIORI ESTIMATES on the solution of a system of differential equations. These estimates provide uniform bounds on a local solution, which allow us to extend this local solution to a global one. For this purpose, it is in general sufficient to establish an estimate of the form ku(t)kX ≤ ϕ(t) ,

t ∈ [0, τ[ ,

(2.53)

where u is a local solution of the system defined on a (maximal) interval [0, τ[, and ϕ : R → R≥0 is continuous and globally defined. For example, estimate (2.18) of proposition 2.7 for the local solutions of Lorenz’ equations is of the form (2.53), with ϕ(t) = ku0 kX eΛt . However, estimates like (2.53) are in general not sufficient to ensure the existence of an attractor. Indeed, if an attractor exists, proposition 2.45 implies that the system is dissipative, that is, the semiflow admits a bounded absorbing set. In particular, the dissipativity of a system forces all solutions to be bounded eventually (once the orbits enter the absorbing set). It follows that the boundedness of solutions is a necessary condition for dissipativity. Therefore, as a first step towards the existence of an attractor, it is desirable to improve estimate (2.53), with a function ϕ which is bounded; that is, to obtain an estimate of the form ku(t)kX ≤ M ,

(2.54)

where M > 0 is independent of t ∈ R≥0 . For example, in proposition 2.65 below, we will see that if σ and b are positive, we can improve (2.18) into a uniform estimate of the form (2.54). In this section we present two results that are commonly used in order to obtain a priori estimates, respectively of the form (2.53) and (2.54). These results are known, respectively, as the G RONWALL’ S INEQUALITY and the EXPONENTIAL INEQUAL ITY .

84

2

Attractors of Semiflows

2.9.1 Integral and Differential Inequalities In this section we prove one version of the G RONWALL’ S for linear differential inequalities.

INEQUALITY ,

and a

COMPARISON THEOREM

PROPOSITION 2.62 (Gronwall’s inequality) Let u, v and w be continuous functions defined on an interval [a, b] ⊂ R, with v nonnegative and w continuously differentiable. If u(t) ≤ w(t) +

Z t

u(s)v(s) ds

(2.55)

a

for all t ∈ [a, b], then also u(t) ≤ w(t) +

Z t

Z exp

a

t

v(s) ds v(τ)w(τ) dτ .

(2.56)

τ

If in addition w is nonnegative and increasing, then Z t u(t) ≤ w(t) exp v(s) ds .

(2.57)

a

PROOF The proof of proposition 2.62 is immediate; we report it for completeness. Let U(t) denote the right side of (2.55). Then U 0 (t) = w0 (t) + u(t)v(t) ≤ w0 (t) +U(t)v(t) .

(2.58)

Rt

Let E(t, τ) := exp( τ v(s) ds). Since ∂ E(t, τ) = −E(t, τ)v(τ) , ∂τ after changing t into τ throughout we can rewrite (2.58) as d (E(t, τ)U(τ)) ≤ w0 (τ)E(t, τ) . dτ We integrate this inequality with respect to τ ∈ [a,t]. Using integration by parts and the fact that E(t,t) = 1, we obtain U(t) ≤ E(t, a)U(a) +

Z t

E(t, τ)w0 (τ) dτ

a

= E(t, a)w(a) + E(t,t)w(t) − E(t, a)w(a) −

Z t ∂ a

∂τ

E(t, τ)w(τ) dτ

Z t

= w(t) +

E(t, τ)v(τ)w(τ) dτ . a

(2.59)

2.9

A Priori Estimates

85

Since u(t) ≤ U(t), (2.56) follows. If w is nonnegative and increasing, we can proceed from (2.59) with Z t Z t ∂ E(t, τ) dτ u(t) ≤ w(t) 1 + E(t, τ)v(τ) dτ = w(t) 1 − a ∂τ a = w(t) (1 − E(t,t) + E(t, a)) = w(t)E(t, a) , from which (2.57) follows. To state our next result, we denote by AC([a, b]; R) the space of all absolutely continuous functions defined on a closed interval [a, b]. Recall that if u ∈ AC([a, b]; R), then u is differentiable almost everywhere in [a, b] (see e.g. theorem A.4). PROPOSITION 2.63 (Linear differential inequality) Let u ∈ AC([a, b]; R), and α, β ∈ L1 (a, b; R). If for almost every t ∈ [a, b], u0 (t) ≤ α(t) + β (t)u(t) , then for all t ∈ [a, b] u(t) ≤ u(a) exp

Z

t

Z t Zt β (s) ds + exp β (s) ds α(τ) dτ . a

a

(2.60)

τ

In particular, if α and β are constant, and β 6= 0, then for t ∈ [a, b] u(t) ≤ e(t −a)β u(a) + αβ e(t −a)β − 1 .

(2.61)

Rt

PROOF Setting E(t, τ) := exp( τ β (s) ds), we have d (E(a,t)u(t)) = E(a,t)u0 (t) − β (t)E(a,t)u(t) ≤ E(a,t)α(t) dt for almost every t ∈ [a, b]. Hence, E(a,t)u(t) − u(a) ≤

Z t

E(a, τ)α(τ) dτ , a

from which, for all t ∈ [a, b], u(t) ≤ E(t, a)u(a) +

Z t

E(t, τ)α(τ) dτ . a

This is (2.60), from which (2.61) follows immediately. Note that if β = 0, the corresponding trivial inequality (for constant α) u(t) ≤ u(a) + α(t − a) follows from (2.61) by letting β → 0 for each fixed t.

86

2

Attractors of Semiflows

2.9.2 Exponential Inequality The bounds provided by proposition 2.62 and 2.63 clearly depend on the interval [a, b]; in particular, on its endpoint b. A more favorable situation is when we can establish a linear EXPONENTIAL INEQUALITY on the solution: in this case, we can obtain a uniform bound independent of t. PROPOSITION 2.64 Let X be a Banach space, S a semiflow on X , and u0 , κ ∈ X . Assume that the function ϕ : [0, ∞[ → R defined by ϕ(t) := kS(t)u0 − κk2 is absolutely continuous. Assume further that there exist positive numbers m, M such that for almost all t ≥ 0, ϕ 0 (t) + m ϕ(t) ≤ M .

(2.62)

√ Let ρ := M/m. Then the closed ball B(κ, ρ) is positively invariant, and for all √ η > 0, each closed ball B(κ, ρ + η) is positively invariant and absorbing. PROOF Proposition 2.63 with α(t) ≡ M, β (t) ≡ −m and a = 0 implies that, for all t ≥ 0, ϕ(t) ≤ e−tm ϕ(0) + ρ(1 − e−tm ) .

(2.63)

In particular, if ϕ(0) ≤ ρ + η, η ≥ 0, we deduce that for t ≥ 0 ϕ(t) ≤ e−tm (ρ + η) + ρ(1 − e−tm ) ≤ ρ + η ; √ that is, B(κ, ρ + η) is positively invariant. To show that when η > 0 this ball is √ also absorbing, let G ⊂ X be bounded. There exists then R > ρ + η such that G ⊆ B(κ, R). Let u0 ∈ G. Then ϕ(0) ≤ R2 , and (2.63) implies ϕ(t) ≤ e−tm R2 + ρ(1 − e−tm ). Thus, ϕ(t) ≤ ρ + η for all t ≥ T , with T = 0 if R2 − ρ ≤ η, and T > 0 defined by T=

1 m

2

ln R η−ρ

√ √ otherwise. Thus, S(t)G ⊆ B(κ, ρ + η) for t ≥ T , and we conclude that B(κ, ρ + η) is absorbing. By way of illustration, we apply this result to prove the existence of an absorbing ball for Lorenz’ equations.

2.9

A Priori Estimates

87

PROPOSITION 2.65 For all positive values of σ and b, the semiflow S defined by Lorenz’ equations (1.50) admits a family of bounded, positively invariant absorbing balls in R3 . Thus, if σ and b are positive, Lorenz’ equations are dissipative. PROOF We choose κ := (0, 0, r + σ ) ∈ R3 , and set u0 := (x0 , y0 , z0 ). Since σ , b > 0, we can estimate (dropping the argument t) d |S(t)u0 − κ|2 = 2(x − 0)x˙ + 2(y − 0)y˙ + 2(z − r − σ )˙z dt = 2x(−σ x + σ y) + 2y(rx − y − xz) + 2(z − r − σ )(xy − bz) = −2σ x2 − 2y2 − 2bz(z − r − σ ) ≤ −2σ x2 − 2y2 − b(z − r − σ )2 + b(r + σ )2 . Thus, setting m := min{2, 2σ , b} and M := b(r + σ )2 , we obtain the differential inequality d |S(t)u0 − κ|2 ≤ −m|S(t)u0 − κ|2 + M . dt This means that the function t 7→ ϕ(t) := |S(t)u0 − κ| satisfies (2.62). Applying proposition 2.64 we conclude that every ball B(κ, R) in R3 , with κ = (0, 0, r + σ ) and p R > M/m, is absorbing and positively invariant for S. In a similar way, it is possible to show that for all positive values of k the solution operator defined by Duffing’s equations (1.53) admits a bounded, positively invariant absorbing set in R2 , even in the nonautonomous case λ 6= 0. We can establish this by means of an analogous a priori estimate on the function E(x, y) := 21 k2 x2 + kxy + y2 + 12 x4 of the solution of (1.53). Note that E is positive definite, and its first three terms are the square of an equivalent norm in R2 , as we immediately see by Schwarz’ inequality. We start by multiplying the second equation in (1.53) by 2y and kx, and adding the resulting identities. Using also the identity kxy˙ =

d d (kxy) − kxy ˙ = (kxy) − ky2 , dt dt

and resorting to weighted Cauchy-Schwarz inequalities of the form Axy ≤ ηy2 +Cη x2 ≤ ηy2 + ηx4 +Cη0 , we arrive at an exponential inequality of the type d E(x, y) + αE(x, y) ≤ C , dt

88

2

Attractors of Semiflows

where C > 0 depends on |λ |. Thus, by proposition 2.64 we deduce that for each λ there exists R > 0 (depending on |λ |), such that the set CR := (x, y) ∈ R2 : 12 k2 x2 + kxy + y2 + 12 x4 ≤ R2 is positively invariant and absorbing. Note that CR is not a ball of R2 .

Chapter 3 Attractors for Semilinear Evolution Equations

In this chapter we apply the results described in chapter 2, in particular theorems 2.46 and 2.50, to show the existence of a global attractor for the semiflows generated by two very simple dissipative evolution equations. These are, respectively, a semilinear version of the heat equation, and of the dissipative wave equation. As such, they are a model of, respectively, a parabolic and of a hyperbolic equation. In chapter 7 we shall apply the same methods to other systems of PDEEs arising from various models in mathematical physics.

3.1 PDEEs as Dynamical Systems 3.1.1 The Model IBV Problems Let Ω ⊂ Rn be a bounded domain with smooth boundary ∂ Ω . We wish to investigate the existence of global attractors, in suitable function spaces, for the semiflows generated by the following two “model” initial-boundary value problems (IBVPs in short) on the cylinder Ω × ]0, +∞[: 1) The IBVP for the semilinear heat equation ut − ∆u + g(u) = f

(3.1)

in Ω × ]0, +∞[, together with the initial and boundary conditions u(0, ·) = u0 u = 0 ∂Ω

in {0} × Ω ,

(3.2)

in ]0, +∞[ × ∂ Ω ;

(3.3)

2) The IBVP for the semilinear dissipative wave equation εutt + ut − ∆u + g(u) = f ,

ε > 0,

(3.4)

in Ω × ]0, +∞[, together with the initial and boundary conditions u(0, ·) = u0 , ut (0, ·) = u1

in {0} × Ω ,

(3.5)

89

90

3

Attractors for Semilinear Evolution Equations

u ∂ Ω = 0

in

]0, +∞[ × ∂ Ω .

(3.6)

CONVENTION 3.1 We refer to problem (3.1) + (3.2) + (3.3) as problem (P) , and to problem (3.4) + (3.5) + (3.6) as problem (Hε ) . We can regard (3.4) as a perturbation of (3.1); or, alternatively, (3.1) as a “reduced version” of (3.4), corresponding to the “limit case” ε = 0. Indeed, problems (P) and (Hε ) can be studied together, in the framework of the so-called SINGULAR PER TURBATION theory (see e.g. Lions, [Lio73]). We shall not explore this issue in any detail, however, except for a brief mention, in section 3.6, of a result on the UPPER SEMICONTINUITY of the global attractors of these problems, as ε → 0. On the other hand, while it is true that the basic well-posedness results for problem (Hε ) do hold for arbitrary ε > 0, we are able to establish most results on the existence of different sorts of attracting sets only when ε is sufficiently small. For specific choices of the nonlinearity g, equations (3.1) and (3.4) describe various models of interest in mathematical physics. In particular, (3.1) is a general model for so-called REACTION - DIFFUSION equations, or C HAFEE -I NFANTE equations; equations of this type were first studied by Chafee and Infante in [CI74]. Equation (3.4) is generally known as a G ORDON TYPE equation; these include, in particular, the so-called SINE -G ORDON and the K LEIN -G ORDON equation of quantum mechanics, corresponding respectively to g(u) = sin u and g(u) = u3 − u. For an extensive review of results on both problems (P) and (Hε ), we refer e.g. to Sell-You, [SY02, scts. 5.1, 5.2]. Problems (P) and (Hε ) are qualitatively very different. This is reflected mainly in the fact that the “space” operator in equation (3.1), i.e. the Laplacian −∆, generates an analytic semigroup, while, in contrast, when (3.4) is transformed into a first order system in a suitable product phase space, the semigroup generated by the corresponding “space” operator is only C0 (see section A.3 for the relevant definitions and properties of semigroups). A major consequence of this difference is that the “parabolic” semigroup is compact for large t, while the “hyperbolic” one is not. Thus, the semiflow generated by (3.1) is asymptotically compact, while the one generated by (3.4) is so only up to a uniformly decaying perturbation. The dissipative effects that force this decay in problem (Hε ) are caused by the term ut , and it will be clear from our arguments that the results we establish for (3.4) do not carry over to the nondissipative case, where the dissipation term ut is not present. In both problems, we assume that f : R × Ω → R; that is, f is independent of u. Later on, we will restrict ourselves to the autonomous case, i.e. when f : Ω → R depends only on the space variable x. For reasons of simplicity, we shall only consider the simple (but nontrivial) model of the quantum mechanics equations, correspond-

3.1

PDEEs as Dynamical Systems

91

ing to the case when n ≤ 3 and g(u) := k(u3 − u) ,

k>0

(3.7)

(in this and the next chapter, we take k = 1). However, most of the results we describe in this and the next chapters can be established, with quite similar methods, for more general nonlinearities g, satisfying suitable growth restrictions. For example, for problem (P) the existence of a global attractor can be established if g satisfies the condition −g(r) ≤0 r |r|→+∞ lim sup

(3.8)

(see e.g. Sell-You, [SY02, sct. 5.1]), while for problem (Hε ) the existence of a global attractor can be established if, in addition to (3.8), g and its antiderivative Z u

G(u) :=

g(v) dv 0

are such that for all r ∈ R, −C1 (r2 + 1) ≤ G(r) ≤ C2 (rg(r) + 1) ,

(3.9)

|g0 (r)| ≤ C3 (1 + |r|ρ ) ,

(3.10)

and, if n ≥ 2,

where C1 , C2 , C3 are positive constants independent of r, 0 < ρ ≤ n−2 2 if n ≥ 3, and ρ > 0 is arbitrary if n = 2. For this result, we refer e.g. to Babin and Vishik, [BV92, ch. 1.8], or Temam, [Tem88, ch. IV.3]). In a sense we explain later, these assumptions on g (which are verified in particular by our choice (3.7); take for example, C1 = C2 = 1 in (3.9) and C3 = 3, ρ = 2 in (3.10), for n = 3) are “not too strong”, and allow us to solve both problems (P) and (Hε ) in an appropriate weak sense, so that we can define corresponding solution operators in a suitable Hilbert space X . Moreover, theorem 2.46 can be applied to the solution operator defined by (3.1), and we can deduce the existence of a global attractor for the semiflow generated by the parabolic problem. In contrast, for the hyperbolic problem (Hε ) the cubic growth of g is critical, with respect to the space dimension n = 3, in the sense that, in this case, the presence of a stronger growth (i.e. with larger exponent) would not allow us to establish a suitable weak solution theory for problem (Hε ). On the other hand, when g is cubic we can proceed to show that, at least if ε is sufficiently small, theorem 2.50 can be applied to the solution operator defined by (3.4), and deduce the existence of a global attractor also for the semiflow generated by the hyperbolic problem (Hε ). Still, the procedure we follow is typical, and in fact almost the same for both problems; indeed, our choice of the nonlinearity (3.7), with n ≤ 3, is motivated only by the goal of providing an outline of the general arguments, in one of the simplest nontrivial settings.

92

3

Attractors for Semilinear Evolution Equations

Examples of “parabolic” equations that can be treated in the same way include reaction-diffusion equations, the Cahn-Hilliard equations, the 2-dimensional NavierStokes equations, the Chafee-Infante equations, etc. Furthermore, the same procedure could be followed for quasilinear parabolic evolution equations of monotone type ut − div ζ (∇u) = f ,

(3.11)

with ζ : Rn → Rn strongly monotone, including the so called p-Laplacian operator n

div ζ (∇u) = ∑ ∂i (|∂i u| p−2 ∂i u) ,

p ≥ 2.

i=1

(This operator is a generalization of the usual Laplace operator in Rn , to which it reduces when p = 2.) In section 6.4 of chapter 6, we will prove the existence of a global attractor for a quasilinear evolution equation of the type (3.11), which models the quasi-stationary Maxwell’s equations in a ferromagnetic medium. Unfortunately, we do not know if this result carries over to the hyperbolic perturbation of (3.11), that is, to an hyperbolic dissipative quasilinear equation of the type ε utt + ut − div ζ (∇u) = f . On the other hand, we can treat far fewer models of “hyperbolic” equations. Among these, the sine-Gordon equation, types of Klein-Gordon equations similar to (3.4), and several types of Kirchoff, von Kármán and Cahn-Hilliard equations, including various models of beam and thin plate equations. We shall present a number of these examples in chapter 6, with the goal of showing that, for all these systems of PDEEs, we can follow a similar procedure, which allows us to apply at least one of theorems 2.46, 2.50, or 2.56, and, consequently, deduce the existence of a global attractor for the semiflows corresponding to each of these systems.

3.1.2 Construction of the Attractors The main steps of the procedure we want to describe, leading to the existence of a global attractor for the semiflows generated by evolution equations like (3.1) or (3.4), can be summarized by the following sequence of incremental results. 1. Solution of the problem. In this step, we transform the problem into an abstract evolution equation in an Hilbert space X , consisting of functions of the space variable. We give a precise definition of what we mean by a solution of the problem in X ; in turn, this allows us to define a corresponding semiflow S on X , generated by the differential equation. Formally, in the parabolic case (3.1), S will be defined by [0, ∞[ × Ω 3 (t, x) 7→ (S(t)u0 )(x) := u(t, x, u0 ) ,

(3.12)

3.1

PDEEs as Dynamical Systems

93

where the right side of (3.12) is the value at the point (t, x) of the solution u of problem (P). In the hyperbolic case (3.4), X is the product of two spaces, each again a space of functions of the space variable, and the semiflow S on X will be formally defined by [0, ∞[ × Ω 3 (t, x) 7→ (S(t)(u0 , u1 ))(x) := (u(t, x, u0 , u1 ), ut (t, x, u0 , u1 )) ,

(3.13)

where u(·, ·, u0 , u1 ) is the solution of problem (Hε ), and ut (·, ·, u0 , u1 ) is its time derivative. In general, the solution operator S will be a semiflow only when the system is autonomous, i.e. when the source term f in the equations is independent of t. As we have indicated, the fact that S is actually a semiflow is a consequence of the well-posedness in the large of the Cauchy problems (P) and (Hε ) in the corresponding spaces X . We carry out this part, first by obtaining local solutions of the problem, and then by establishing a priori estimates which allow us to extend the local solutions to global ones. 2. Absorbing sets. In this step we show the existence of a bounded, positively invariant absorbing set B for S in X . As we have seen in the finite dimensional examples of chapter 2, this part can be carried out by refining the a priori estimates established in Step 1. Following the ideas of proposition 2.65 on the absorbing set of Lorenz’ equations, the most effective way to obtain an absorbing set is to establish a linear differential inequality on the square of a norm of u, as in (2.62). In a certain sense, the possibility of doing so is characteristic of dissipative systems; for example, differential inequalities like (2.62) are generally not available for nondissipative wave equations of the form utt − ∆u = F(t, x, u) . 3. Compactness of the semiflow. In this step we establish suitable regularizing properties of the semiflow S. In the parabolic case, this step usually exploits the smoothing effect of parabolic operators, which allows us to deduce directly that S is uniformly compact in X for large t. In the hyperbolic case, we show instead that, at least if ε is sufficiently small, S admits a decomposition S = S1 + S2 as in proposition 2.48, with S1 uniformly compact for large t and S2 uniformly decaying to 0. This will indeed be possible for equation (3.4); however, there are other types of equations, such as the beam equation, the von Kármán equation, or the perturbed nonviscous Cahn-Hilliard equation, for which we do not know how to implement this step. In this case, we would resort to the method of α-contractions, presented in section 2.7. 4. Conclusion. We apply theorems 2.46 or 2.50 (or theorem 2.56 for α-contractions), to deduce the existence of a global attractor A for S in X . This attractor gives a description of the long-time behavior of the solutions of the equations, independently of their initial values. Of course, these steps can be supplemented by others, concerning further properties of these attractors, such as their regularity, finite dimensionality, geometrical, topological or differential structure, etc. For instance, regularity is often established

94

3

Attractors for Semilinear Evolution Equations

by proving that A is contained in a subspace X1 ,→ X . If the injection X1 ,→ X is compact, compactness of A in X can also be obtained by showing that A is bounded in X1 . As an illustration, in section 3.5 we shall show a regularity result of this type for problem (Hε ). REMARK 3.2 In sections 3.3 and 3.4 we shall present a direct implementation of step 1, for both problems (P) and (Hε ). That is, we shall give a constructive procedure, based on an approximation technique, which yields the solutions to both problems, and allows us to generate the corresponding semiflows. An alternative approach, which is more in analogy to the classical methods of ODEs, is provided by the theory of SEMIGROUPS (see section A.3). In this setting, we transform the problem into an abstract evolution equation of the form Ut + AU = F(U) ,

(3.14)

in an Hilbert space X , consisting of functions of the space variable. In fact, problem (P) is already in the form (3.14), while for problem (Hε ), X is a product space, and A is a formal matrix, acting on the vector U := (u, ut )> , in accord with (3.13). When both problems are transformed into the form (3.14), the linear operator A is densely defined, with domain dom(A) := {u ∈ X : Au ∈ X } , and generates at least a C0 -SEMIGROUP (e−tA )t ≥0 on X . This allows us to transform (3.14) into an integral equation, and, in analogy to the classical procedure in ODEs, consider the corresponding solutions of the initial value problem. In particular, we have the following DEFINITION 3.3 Let U0 ∈ X , and T > 0. A function U ∈ C([0, T ]; X ) is a MILD SOLUTION of equation (3.14), with initial value U(0) = U0 , if U satisfies the integral equation U(t) = e−tAU0 +

Z t

e−(t −θ )A F(U(θ )) dθ ,

0≤t ≤T .

(3.15)

0

Note that (3.15) is a straightforward generalization of the familiar Duhamel’s formula for the solution of the system of ODEs u˙ + Au = F(u) , in which A is a constant matrix in RN . The following result (see theorem A.52, whose statement we repeat here for convenience) shows that the initial value problem for (3.14) is well posed in the class of mild solutions. THEOREM 3.4 Assume that F : X → X is globally Lipschitz continuous. Then for all U0 ∈ X , the initial value problem for (3.14), with initial value U(0) = U0 , has a unique mild

3.2

Functional Framework

95

solution u. Moreover, for all T > 0 the map X 3 U0 7→ U ∈ C([0, T ]; X ) is Lipschitz continuous. In particular, theorem 3.4 implies that the evolution equation (3.14) generates a continuous semiflow S on X . We shall refer to this approach again in section 5.6.1 of chapter 5, when we construct inertial manifolds for general evolution equations of the form (3.14). For a comprehensive introduction to semigroup theory, we refer e.g. Pazy, [Paz83], or Engel and Nagel, [EN00]; we report the basic results we need in section A.3.

3.2 Functional Framework In this section we introduce the functional space framework in which we study problems (P) and (Hε ). We also recall some well known facts on the Laplace operator ∆, associated to homogeneous Dirichlet boundary conditions. Throughout the rest of this chapter, we assume that Ω is a bounded domain of Rn , with a smooth (i.e. at least Lipschitz) boundary ∂ Ω .

3.2.1 Function Spaces We consider the following spaces and adopt the following notations: 1. H := L2 (Ω ), with norm k · k and scalar product h· ,·i. 2. V := H10 (Ω ), with norm kukV = k∇uk; this is justified by Poincaré’s inequality p ∃ λ1 > 0 ∀ u ∈ V : λ1 kuk ≤ k∇uk (3.16) (see (A.73)). The corresponding scalar product in V is then hu, viV := h∇u, ∇vi . 3. V 0 := H−1 (Ω ), with norm k · kV 0 . V 0 is the topological dual of V. We denote the duality pairing between V 0 and V by (·, ·)V 0 ×V . 4. D := H2 (Ω ) ∩ H10 (Ω ). Since the Poincaré type inequality p λ1 k∇uk ≤ k∆uk

(3.17)

holds for all u ∈ D, we can, and will, choose in D the norm kukD := k∆uk. To prove (3.17), note that if u ∈ D, then k∇uk2 = h−∆u, ui ≤ k∆uk kuk , so (3.17) follows by (3.16).

96

3

Attractors for Semilinear Evolution Equations

5. L : H → H is the unbounded operator formally defined by Lu = −∆u, with domain dom(L) = D . ∆ is the usual Laplace operator, formally defined by n

∆ :=

∂2

∑ ∂ x2 .

j=1

j

Note that, in choosing D as the domain of L, we are automatically imposing homogeneous Dirichlet boundary conditions on ∂ Ω , in the weak sense of H10 (Ω ). 6. For 1 ≤ p ≤ +∞, |·| p denotes the norm in L p (Ω ), and q ∈ [1, +∞] denotes the conjugate index of p, i.e. q := 1 if p = +∞, q := +∞ if p = 1, and q := p−p 1 otherwise. 7. If X is a function space on Ω and f : [0, T ] → X , we implicitly define a function on [0, T ] × Ω , with value f (t)(x) for (t, x) ∈ [0, T ] × Ω . We denote this function again by f ; that is, we set f (t, x) := f (t)(x) ,

(t, x) ∈ [0, T ] × Ω .

Similarly, given a function g : [0, T ] × Ω → R, we define a function from [0, T ] into X , with value g(t, ·) at t ∈ [0, T ]. We denote this function again by g; that is, we set g(t)(x) := g(t, x) , (t, x) ∈ [0, T ] × Ω . Likewise, we denote by ∆ f (t) (respectively, ∇ f (t)) the function with value ∆ f (t, x) (respectively, ∇ f (t, x)) at x ∈ Ω . 8. Finally, for T > 0 we set W(T ) := { f ∈ L2 (0, T ; V) : ft ∈ L2 (0, T ; V 0 )} , and recall that W(T ) ,→ C([0, T ]; H) .

(3.18)

Imbedding (3.18) is part of the following result, whose proof can be found e.g. in Tanabe, [Tan79] (Lemma 5.5.1): PROPOSITION 3.5 Let V and H be separable Hilbert spaces, with V ,→ H continuously and densely. If u ∈ W(T ), u can be modified on a set of measure 0, in such a way that u ∈ C([0, T ]; H). If also v ∈ W(T ), the function t 7→ hu(t), v(t)iH is absolutely continuous on [0, T ], and satisfies the identity d hu(t), v(t)iH = (ut (t), v(t))V 0 ×V + (vt (t), u(t))V 0 ×V dt

3.2

Functional Framework

97

in D0 (]0, T [) and, in fact, for almost every t ∈ [0, T ]. In particular, taking v = u we deduce that for almost all t ∈ [0, T ], d ku(t)k2H = 2 (ut (t), u(t))V 0 ×V . dt

3.2.2 Orthogonal Bases As we recall in theorem A.76, the eigenvalue problem for L ( −∆w j = λ j w j w j |∂ Ω = 0

(3.19)

admits an unbounded sequence of positive eigenvalues (λ j ) j∈N . This sequence can be ordered so that 0 < λ1 ≤ λ2 ≤ · · · ≤ λ j ≤ · · · ,

λ j → +∞ .

(3.20)

For each j ∈ N, the corresponding eigenvector w j is in C∞ (Ω ) ∩ C(Ω ), and the sequence (w j ) j∈N is a complete orthogonal system in L2 (Ω ). This means that each u ∈ L2 (Ω ) has a uniquely determined Fourier series expansion with respect to the w j ’s; more precisely, ∞

u=

∑ hu, w j iw j ,

(3.21)

j=1

with the series (3.21) converging in L2 (Ω ). Since rescaled eigenvectors are still eigenvectors, we can assume that the (w j ) j∈N are in fact an orthonormal system, i.e. that kw j k = 1 for all j ∈ N. Then, Parseval’s formula holds: for all u ∈ H, ∞

kuk2 =

∞

∑ hu, w j i2 kw j k2 = ∑ hu, w j i2 .

j=1

(3.22)

j=1

The particular choice (3.19) of eigenvectors assures the orthogonality (but, of course, not the orthonormality) of the sequence (w j ) j∈N also in V. In fact, for arbitrary j and k we compute that h∇w j , ∇wk i = h−∆w j , wk i = λ j hw j , wk i ;

(3.23)

thus, if j 6= k, h∇w j , ∇wk i = 0 , while if j = k h∇w j , ∇w j i = k∇w j k2 = λ j kw j k2 = λ j .

(3.24)

As a consequence, in analogy to (3.22) we have that for all u ∈ V, ∞

k∇uk2 =

∞

∑ hu, w j i2 k∇w j k2 = ∑ λ j hu, w j i2 .

j=1

j=1

(3.25)

98

3

Attractors for Semilinear Evolution Equations

3.2.3 Finite Dimensional Subspaces We will construct solutions to problems (P) and (Hε ) as limits of sequences of approximating solutions, in the framework of a suitable Galerkin scheme. In the implementation of this type of scheme, the starting step is the projection of the problem into a sequence of finite dimensional subspaces of V, on each of which the PDEE is reduced to a finite system of ODEs. We will recall this procedure when we consider problem (P). This step requires the choice of a so-called TOTAL BASIS of V. As we recall in definition A.10, this is a countable set of linearly independent vectors w j , such that ∞ [

Vm = V

m=1

(closure in V), where for m ∈ N, Vm := span{w1 , . . . , wm } .

(3.26)

Total bases of V exist, because V is separable. Among these, a particularly convenient one is the sequence of eigenvectors (w j ) j∈N of L, defined in (3.19). Referring then to the series expansion (3.21), for each N ∈ N we define a projection PN : V → VN by setting P0 := 0 and, if N ≥ 1, by ∞

N

PN (u) :=

∑ hu, w j iw j

if u =

j=1

∑ hu, w j iw j .

(3.27)

j=1

We also set QN := I − PN ; that is, QN (u) is the tail of the series (3.21). Both PN and QN are clearly orthogonal projections in H; because of (3.23), they are also orthogonal in V. It is also worth recalling that the family (PN )n≥1 is monotone; that is, if n ≤ m, then Pn ≤ Pm . This inequality is to be understood in the operator sense, that is, that hPn x, xi ≤ hPm x, xi for all x ∈ X . To see this, it is sufficient to note that, because of the orthogonality of the sequence (w j ) j∈N , from (3.21) and (3.27) we have n

hPn x, xi =

m

∑ hx, w j i2 ≤ ∑ hx, w j i2 = hPm x, xi .

j=1

j=1

The following result, which generalizes Poincaré’s inequality (3.16), will be used in chapters 4 and 5: PROPOSITION 3.6 For all N ∈ N and u ∈ V, k∇(QN (u))k2 ≥ λN+1 kQN (u)k2 ,

(3.28)

3.3

99

The Parabolic Problem

k∇(PN (u))k2 ≤ λN kPN (u)k2 .

(3.29)

∞

PROOF Let u = ∑ hu, w j iw j , as in (3.21). Then (3.28) follows from (3.22), (3.25) j=0

and (3.20), since ∞

k∇(QN (u))k2 =

∑

∞

λ j hu, w j i2 ≥ λN+1

j=N+1

∑

hu, w j i2 = λN+1 kQN (u)k2 .

j=N+1

This proves (3.28); the proof of (3.29) is similar. Note that (3.16) corresponds to (3.28) with N = 0.

3.3 The Parabolic Problem In this section we illustrate how to implement the steps we have listed in section 3.1.2 for the model parabolic IBVP (3.1). We choose X := H = L2 (Ω ), and set W0 (T ) := { f ∈ W(T ) : f (T, ·) = 0} ; note that W0 (T ) is well defined, because by proposition 3.5 the trace f (T, ·) makes sense at least in X . We need the following generalization of proposition 3.5: PROPOSITION 3.7 Let V and H be as in proposition 3.5, and assume that u ∈ L2 (0, T ; V) ∩ L p (]0, T [ × Ω ) =: Y , ut ∈ L2 (0, T ; V 0 ) + Lq (]0, T [ × Ω ) =: Z ,

(3.30)

for some p ∈ ]1, +∞[, with q = p−p 1 . Then u can be modified on a set of measure 0 so that u ∈ C([0, T ]; H). Moreover, the function t 7→ ku(t)k2H is absolutely continuous on [0, T ], and if ut = g1 + g2 , with g1 ∈ L2 (0, T ; V 0 ) and g2 ∈ Lq (]0, T [ × Ω ) (which is the meaning of the second of (3.30)), then for almost all t ∈ [0, T ], d ku(t)k2H = 2 (g1 (t), u(t))V 0 ×V + 2hg2 (t), u(t)iH . dt

(3.31)

PROOF We briefly sketch the proof of proposition 3.7 for the sake of completeness. We recall that Z = Y 0 , via the duality Z T

(g1 + g2 , u)Z×Y :=

0

Z T

(g1 (t), u(t))V 0 ×V dt +

0

hg2 (t), u(t)iH dt ;

100

3

Attractors for Semilinear Evolution Equations

note that the product g2 u belongs to L1 (]0, T [ ×Ω ). If u ∈ C1 ([0, T ]; H), from the identity d ku(t)k2H = 2hut (t), u(t)iH = 2 (g1 (t) + g2 (t), u(t))Y 0 ×Y dt = 2 (g1 (t), u(t))V 0 ×V + 2hg2 (t), u(t)iH ,

(3.32)

we deduce that for 0 ≤ s < t ≤ T ku(t)k2H = ku(s)k2H + 2

Z t s

Z t

(g1 (θ ), u(θ ))V 0 ×V dθ + 2

s

hg2 (θ ), u(θ )iH dθ . (3.33)

Choosing s so that ku(s)k2H =

1 T

Z T 0

ku(θ )k2H dθ ,

we deduce from (3.33) that 1 ku(t)kH ≤ T 2

Z T 0

Z T

+2 0

2

ku(θ )kH dθ + 2

Z T 0

kg1 (θ )kV 0 ku(θ )kV dθ

|g2 (θ )|q |u(t)| p dθ .

From this, we obtain that kukC([0,T ];H) ≤ T1 kuk2L2 (0,T ;H) + 2kut kZ kukY . We can then easily conclude that {u ∈ Y : ut ∈ Y 0 } ,→ C([0, T ]; H) by a density argument. Likewise, (3.31) follows from (3.32), noting that its right side is in L1 (0, T ).

3.3.1 Step 1: The Solution Operator We now prove a global existence and well-posedness result for problem (P), which allows us to define the associated solution operator. If the equation is autonomous, i.e. if f only depends on x, this solution operator is in fact a continuous semiflow. DEFINITION 3.8 Let T > 0, u0 ∈ H, f ∈ L2 (0, T ; V 0 ). A function u is a SOLUTION of problem (P), with g(u) = u3 − u, on the interval [0, T ] if

WEAK

u ∈ C([0, T ]; H) ∩ L2 (0, T ; V) ∩ L4 (]0, T [ × Ω ), and for all ϕ ∈ W0 (T ) ∩ L4 (]0, T [ × Ω ), Z T 0

Z − hu, ϕt i + h∇u, ∇ϕi + hu3 − u, ϕi dt = 0

T

( f , ϕ)V 0 ×V dt + hu0 , ϕ(0)i . (3.34)

3.3

The Parabolic Problem

101

Note that the left side of (3.34) makes sense: in fact, by theoremA.58, the product u3 ϕ is integrable if u ∈ L4 (]0, T [ × Ω ) and ϕ ∈ L4 (]0, T [ × Ω ). Weak solvability of problem (P) is then assured by THEOREM 3.9 For all T > 0, u0 ∈ H and f ∈ L2 (0, T ; V 0 ), there exists a unique weak solution u of problem (P). This solution depends continuously on the initial value u0 , and the function t 7→ ku(t)kH is absolutely continuous on [0, T ]. If in addition u0 ∈ V and f ∈ L2 (0, T ; H), then u ∈ C([0, T ]; V) ∩ L2 (0, T ; H2 (Ω )) , ut ∈ L2 (0, T ; H) ,

(3.35)

and the function t 7→ ku(t)kV is absolutely continuous on [0, T ]. PROOF 1. We first remark that if problem (P) does admit a weak solution u as claimed, then (3.34) implies that the identity ut = f + u − u3 + ∆u

(3.36)

holds at least in D0 (0, T ; V 0 ). Thus, reading (3.36) as ut = ( f + u + ∆u) − u3 =: g1 − g2 ∈ L2 (0, T ; V 0 ) + L4/3 (]0, T [ × Ω ) , by proposition 3.7 we deduce that u ∈ C([0, T ]; H), and that the function t 7→ ku(t)kH is absolutely continuous. Finally, if u enjoys the additional regularity described in (3.35), then proposition 3.7, applied to each partial derivative ∂k u, 1 ≤ k ≤ n, with p = q = 2, implies that the function t 7→ ku(t)kV is also absolutely continuous. 2. A proof of the existence and regularity part of theorem 3.9 can be given by means of a standard Galerkin approximation method, as e.g. in Lions, [Lio69], or Temam, [Tem88, sct. III.1.1]. We sketch the details of the proof of the existence claim for completeness. We refer to the finite dimensional subspaces Vm of V introduced in section 3.2.3 and, for each m ∈ N, project problem (P) on Vm . That is, we look for a function (t, x) 7→ um (t, x) (here, um does not mean u to the power m), with um (t, ·) ∈ Vm for t ∈ [0, T ], determined as the solution of the system of m ODEs ( hutm (t) − ∆um (t) + (um (t))3 − um (t) − f (t), w j i = 0 , (3.37) j = 1, . . . , m , supplemented by the initial conditions um (0) = um 0 := Pm (u0 ) ∈ Vm .

(3.38)

Thus, um should have the form m

um (t, x) =

∑ γ j (t)w j (x) ,

j=1

(3.39)

102

3

Attractors for Semilinear Evolution Equations

and (3.37) is a system of ODEs for the scalar components γ1 , . . . , γm of um ((3.39) means that we solve the projected PDE by separation of variables). The Cauchy problem (3.37)+(3.38) can be solved by means of Carathéodory’s theorem A.5, which provides a local solution um ∈ AC([0,tm ]; Vm ) of (3.37), defined on some interval [0,tm ] ⊆ [0, T ] (that is, each γ j ∈ AC([0,tm ])). We then extend each um to all of [0, T ], by means of the a priori estimates that follow. Before presenting these estimates, we introduce two conventions, that we follow not only here, but almost always in the remainder of these notes. CONVENTION 3.10 In writing estimates, we often omit, for convenience, one of the variables t and x, or both. This means that, for example, if u : [0, T ] → L2 (Ω ), we write kuk instead of ku(t, ·)k. CONVENTION 3.11 Unless otherwise specified, we denote by C a generic constant, which may be different from line to line of the same estimate, or even within the same line. This means that, for example, we identify quantities like C, C2 , eC , etc. On the other hand, when a constant depends on a parameter in a crucial way (e.g., if the constant is unbounded as the parameter vanishes), we shall always indicate this dependence explicitly. We proceed then to multiply each equation (3.37) by γ j (t). Summing all the resulting identities for j = 1, . . . , m, we obtain d m ku k + 2k∇um k2 + 2|um |44 = 2h f + um , um i ≤ k f k2V 0 + k∇um k2 + 2kum k2 . dt We integrate this inequality on [0,t], 0 < t ≤ tm : kum (t)k2 +

Z t

k∇um (θ )k2 dθ + 2

Z t

0

≤

0

2 kum 0k +

Z T 0

|um (θ )|44 dθ

2

k f (θ )kV 0 dθ + 2

Z t

kum (θ )k2 dθ .

(3.40)

0

Recalling that the sequence (um 0 )m∈N is bounded in H (because it converges to u0 ), by Gronwall’s inequality (proposition 2.62) we derive from (3.40) the estimate kum kL∞ (0,tm ;H) + kum kL2 (0,tm ;V ) + kum k2L4 (]0,tm [ ×Ω ) ≤ M1 etm , where 2 M12 := sup kum 0k + m≥1

Z T 0

k f (θ )k2V 0 dθ

is finite. Since M1 etm ≤ M1 eT =: M ,

(3.41)

3.3

103

The Parabolic Problem

we obtain from (3.41) an estimate independent of both tm and m (however, M depends on T , as is typical of a priori estimates obtained by means of Gronwall’s inequality). The fact that M is independent of tm allows us to extend um to all of [0, T ], by a usual continuation argument (see theorem A.2). The fact that M is independent of m means that the sequence (um )m∈N is bounded in each of the spaces L∞ (0, T ; H) , L2 (0, T ; V) , L4 (]0, T [ × Ω ) .

(3.42)

Hence, by theorem A.16, there is a subsequence of (um )m∈N , which we still denote by (um )m∈N , converging weakly to a limit u in each of the spaces of (3.42). Moreover, since the sequence ((um )3 )m∈N is bounded in L4/3 (]0, T [ × Ω ), there is a further subsequence, which we again keep denoting by (um )m∈N , such that (um )3 converges to a limit χ weakly in L4/3 (]0, T [ × Ω ). Fixing j in (3.37) and letting m → +∞, we deduce that u and χ solve the equation ut − ∆u + χ − u = f . Thus, to complete the proof of the existence part of theorem 3.9 it is sufficient to show that χ = u3 . To this end, we fist show that um → u

in L2 (0, T ; H) strongly .

(3.43)

To obtain this, we recall from theorem A.74 that the Sobolev imbedding H1 (Ω ) ,→ L4 (Ω ) holds, since n ≤ 3. In turn, this implies that L4 (0, T ; V) ,→ L4 (]0, T [ × Ω ) and, therefore, that the dual imbedding L4/3 (]0, T [ × Ω ) ,→ L4/3 (0, T ; V 0 ) holds. From this it follows that the sequence ((um )3 )m∈N is also bounded in the space L4/3 (0, T ; V 0 ). In fact, we have that k(um )3 kL4/3 (0,T ;V 0 ) ≤ Ck(um )3 kL4/3 (]0,T [ × Ω ) = Ckum k3L4 (]0,T [ × Ω ) , and this last sequence is bounded, since (um )m∈N is bounded in L4 (]0, T [ × Ω ). Now, (3.37) implies that utm = Pm f + um − (um )3 + ∆um , and, therefore, that also the sequence (utm )m∈N is bounded in L4/3 (0, T ; V 0 ). From theorem A.82, we have that the injection {u ∈ L2 (0, T ; V) : ut ∈ L4/3 (0, T ; V 0 )} ,→ L2 (0, T ; H) is compact; hence, (3.43) follows. As a consequence of (3.43), (um )3 → u3 in L1 (]0, T [ × Ω ) strongly. This follows from the estimate k(um )3 − u3 kL1 (]0,T [ × Ω ) ≤ k(um )2 + um u + u2 kL2 (]0,T [ × Ω ) kum − ukL2 (]0,T [ × Ω ) ,

104

3

Attractors for Semilinear Evolution Equations

and from the fact that the sequence ((um )2 )m∈N is bounded in L2 (]0, T [ × Ω ), since k(um )2 kL2 (]0,T [ × Ω ) = kum k2L4 (]0,T [ × Ω ) . Therefore, for all ψ ∈ L∞ (]0, T [ × Ω ), Z TZ

(um (t, x))3 ψ(t, x) dx dt →

Im := 0

Z TZ 0

Ω

(u(t, x))3 ψ(t, x) dx dt .

(3.44)

Ω

On the other hand, since evidently ψ ∈ L4 (]0, T [ × Ω ), and (um )3 → χ weakly in the dual space L4/3 (]0, T [ × Ω ), Im →

Z TZ

χ(t, x) ψ(t, x) dx dt . 0

(3.45)

Ω

From (3.44) and (3.45) we deduce that for each ψ ∈ L∞ (]0, T [ × Ω ), Z TZ 0

(u(t, x))3 ψ(t, x) dx dt =

Z TZ

χ(t, x) ψ(t, x) dx dt . 0

Ω

Ω

Since L∞ (]0, T [ × Ω ) is isomorphic to the dual of L1 (]0, T [ × Ω ), by the HahnBanach theorem (see theorem A.13), we conclude that χ = u3 in L1 (]0, T [ × Ω ). This completes the proof of the existence part of theorem 3.9. 3. To show the well-posedness of problem (P), we consider the difference z = u − v of two solutions of (3.1), which solves the equation zt = z + ∆z − (u3 − v3 ) .

(3.46)

Since z + ∆z ∈ L2 (0, T ; V 0 ) , u3 − v3 ∈ L4/3 (]0, T [ × Ω ) , by proposition 3.7 we obtain d kzk2 = 2 (z + ∆z, z) − 2hu3 − v3 , u − vi dt almost everywhere in t. Since 2hu3 − v3 , u − vi ≥ 0 (by monotonicity), we obtain that, for almost all t, d kz(t)k2 ≤ 2kz(t)k2 − 2k∇z(t)k2 ≤ 2kz(t)k2 . dt

(3.47)

Applying the comparison theorem (proposition 2.63), we deduce from (3.47) that kz(t)k2 ≤ kz(0)k2 e2t , that is, ku(t) − v(t)k ≤ ku(0) − v(0)k et .

(3.48)

3.3

105

The Parabolic Problem

Thus, uniqueness and well-posedness in the large follow. We can now define the solution operator generated by problem (P). Indeed, since T is arbitrary in theorem 3.9, we have that if f ∈ L2loc (0, +∞; V 0 ), then u ∈ C([0, +∞[; H); thus, theorem 3.9 defines a unique global solution of problem (P). If f is independent of t, i.e. if f (t) ≡ f ∈ V 0 , theorem 3.9 defines, by means of (3.12), a continuous semiflow S on X = H. This is the semiflow generated by problem (P). In particular, note that inequality (3.48) shows that each operator S(t), t ≥ 0, is Lipschitz continuous in X . Finally, we would like to mention a possible modification of the setting of problem (P), motivated by the fact that in order to proceed with the construction of the attractor for the semiflow S so defined, we shall need to assume that f (t) ≡ f ∈ H (as opposed to V 0 ). Thus, in theorem 3.9 we could assume that f ∈ L2 (0, T ; H); as we now show, this would allow us to simplify the choice of the space of test functions W0 (T ) into ˜ 0 (T ) := { f ∈ L2 (0, T ; V) : ft ∈ L2 (0, T ; H), f (T, ·) = 0} . W

(3.49)

In fact, in this case we have the imbeddings ˜ 0 (T ) ,→ C([0, T ]; H1/2 (Ω )) , W

H1/2 (Ω ) ,→ L3 (Ω ) ,

the second of which holds because n ≤ 3 (see theorems A.81 and A.74). Thus, we ˜ 0 (T ) ,→ C([0, T ]; L3 (Ω )). By means of the interpolation inequality deduce that W 1/2

1/2

|ϕ|4 ≤ C|ϕ|6 |ϕ|3 , ˜ 0 (T ) ,→ L4 (]0, T [ × Ω ) as well. Consequently, (see theorem A.61), we deduce that W 3 ˜ 0 (T ), the product u ϕ is again in L1 (]0, T [ × Ω ), and equation (3.34) makes if ϕ ∈ W sense.

3.3.2 Step 2: Absorbing Sets In this step we prove the existence of a bounded absorbing set B for S in X , by establishing suitable a priori estimates on u in X (recall that X = H = L2 (Ω )). PROPOSITION 3.12 Assume f ∈ Cb ([0, +∞[; X ). There exist positive constants m1 and λ1 , such that for all u0 ∈ X and t ≥ 0, m1 m1 e−2λ1 t + 2λ . (3.50) ku(t)k2 ≤ ku0 k2 − 2λ 1

1

Consequently, there exists R > 0 such that the ball B(0, R) in X is absorbing and positively invariant for S.

106

3

Attractors for Semilinear Evolution Equations

PROOF Formally multiplying equation (3.1) by 2u (as in the proof of uniqueness in theorem 3.9, this can be justified by means of proposition 3.5), we obtain, dropping the argument t for convenience: d kuk2 + 2k∇uk2 + 2|u|44 = 2h f + u, ui ≤ C1 + k f k2 + |u|44 , dt

(3.51)

where C1 is a constant depending only on Ω . Recalling Poincaré’s inequality (3.16), we deduce from (3.51) d kuk2 + 2λ1 kuk2 ≤ C1 + sup k f (t)k2 =: m1 . dt t ≥0

(3.52)

Inequality (3.52) is a linear differential inequality, like (2.62); since the function t 7→ ku(t)k2 is absolutely continuous, estimate p (3.50) follows from (3.52), via proposition 2.64. We deduce then that for any R > m1 /(2λ1 ), the ball B(0, R) in X is absorbing and positively invariant for S. In particular, we find that S(t)u0 ∈ B(0, R) for all t ≥ T0 , with T0 = 0 if ku0 k ≤ R or, if ku0 k > R, T0 =

m1 ku0 k2 − 2λ 1 ln 2 m1 1 . 2λ1 R − 2λ 1

This concludes the proof of proposition 3.12; we mention that in chapter 4, section 4.3.1, we shall show the existence of a positively invariant ball B1 , bounded and absorbing in V.

3.3.3 Step 3: Compactness of the Solution Operator We now establish the uniform compactness of the solution operator S for large t, by means of further a priori estimates on the solution of (3.1). These estimates show that u(t) ∈ V if t is sufficiently large. We assume again that f ∈ Cb ([0, +∞[; X ). In the sequel, we denote by C, C1 , C2 , . . . any generic positive constant, possibly depending on f or Ω , but not on u (recall convention 3.11). We formally multiply (3.1) in X by −(et − 1)∆u(t) and obtain (et − 1)h∇ut (t), ∇u(t)i + (et − 1)k∆u(t)k2 − (et − 1)hu3 (t), ∆u(t)i = −(et − 1)h f (t) + u(t), ∆u(t)i .

(3.53)

This procedure is formal, because we do not know that ∆u(t) ∈ L2 (Ω ); in fact, this is even more than what we are trying to establish. To proceed rigorously, we should establish the estimates that follow for the Galerkin approximations of u, and then realize that the final estimate we obtain can be carried over to u itself. Proceeding from (3.53), we obtain d (et − 1)k∇u(t)k2 + 2(et − 1)k∆u(t)k2 + 6(et − 1)hu2 (t)∇u(t), ∇u(t)i dt

3.3

107

The Parabolic Problem

= et k∇u(t)k2 − 2(et − 1)h f (t) + u(t), ∆u(t)i = k∇u(t)k2 − (et − 1)h2 f (t) + 3u(t), ∆u(t)i ≤ k∇u(t)k2 +C2 (et − 1) k f (t)k2 + ku(t)k2 + (et − 1)k∆u(t)k2 . From this, estimating ku(t)k by (3.50), we obtain d (et − 1)k∇u(t)k2 + (et − 1)k∆u(t)k2 dt

≤ k∇u(t)k2 +C2 (et − 1) m1 + ku0 k2 + mλ 1 1

.

(3.54)

Integration of (3.51) yields, in particular, that Z t

2 0

k∇u(s)k2 ds ≤ ku0 k2 + m1t ;

thus, from (3.54) we obtain (neglecting a positive term at its left side) m1 =: C3 et . (et − 1)k∇u(t)k2 ≤ C2 et m1 + ku0 k2 + 2λ 1

From this we deduce that, if e.g. t > ln 2, k∇u(t)k2 ≤

C3 et ≤ 2C3 . et − 1

This proves the asserted uniform compactness of the solution operator S for large t. For future reference, we remark that if u0 ∈ V, the same estimates (in fact, simpler, because we do not need the factor et − 1) would yield the regularity result u ∈ Cb ([0, +∞[; V) ∩ L2 (0, +∞; H2 (Ω )) for the solution of problem (P).

3.3.4 Step 4: Conclusion If problem (P) is autonomous, we can now deduce the existence of the global attractor for S, as a consequence of theorem 2.46. In conclusion, we have: THEOREM 3.13 Let u0 , f ∈ L2 (Ω ). The initial boundary value problem (P) defines a semiflow S in X = L2 (Ω ), which admits a global attractor in X . This attractor is the set A := ω(B) =

\[

S(t)B

s≥0 t ≥s

(closure in X ), where B is the absorbing ball B(0, R) determined in proposition 3.12. Much more information is available on the properties of the global attractor A of the semiflow S generated by problem (P) obtained in theorem 3.13. Among these, we mention the following.

108

3

Attractors for Semilinear Evolution Equations

1. Structure of the Attractor. In addition to all stationary solutions of problem (P), that is, the solutions of the nonlinear elliptic boundary value problem ( −∆u + u3 − u = f , (3.55) u|∂ Ω = 0 , A also contains the unstable manifolds (in X ) of all these stationary solutions (see e.g. Babin-Vishik, [BV92, ch. 3]). 2. Attractors via α-contractions. The existence of A can also be established by means of theorems 2.56 and 2.55. Indeed, as we have already remarked, since S is compact for large t, the operator T = S(t∗ ) is an α-contraction if t∗ is large enough. 3. Regularity of the Attractor. The regularizing effect of the heat operator would allow us to show that A is in fact contained and bounded in V ∩ H2 (Ω ); thus, A is compact not only in X , but also in V. For a proof of this result on the regularity of the attractor, see e.g. Temam, [Tem88, ch. 3], or Babin-Vishik, [BV92, sct. 3.3]. 4. Finite dimensionality of the Attractor. The fractal dimension dimF (A) of A is finite. While this can be proven directly (see e.g. Temam, [Tem88, sct. VI.2]), we give an alternative proof of this result in (4.87) of section 4.5.6 in the next chapter, where we prove the existence of an exponential attractor E for S. The finite dimensionality of A is in fact a consequence of that of E, and of the fact that A ⊆ E, which implies that dimF (A) ≤ dimF (E).

3.3.5 Backward Uniqueness The heat equation cannot be solved backward in time directly. Indeed, even in the autonomous case (i.e., with f independent of t), the usual change of variable v(t, x) := u(−t, x), t < 0, transforms the heat equation (3.1) into the equation vt + ∆v = f + g(v) , in which +∆ is a negative operator in L2 (Ω ). Consequently, problem (P) actually defines a semiflow only. Nevertheless, we can extend S to a flow, by means of proposition 2.5. This requires that each operator S(t), t > 0, be invertible; in turn, this requires the semiflow generated by the heat equation to satisfy a BACKWARD UNIQUE NESS property, in the sense of the following DEFINITION 3.14 The semiflow S satisfies the BACKWARD UNIQUENESS PROP ERTY if whenever x, y ∈ X and t > 0 are such that S(t)x = S(t)y, then x = y. To this end, we first need the following property of the nonlinear term of the equation:

3.3

109

The Parabolic Problem

PROPOSITION 3.15 The function u 7→ u3 is locally Lipschitz continuous from V into H. PROOF First we note that u3 ∈ H if u ∈ V. In fact, Z

|u3 |2 dx =

Ω

Z

|u|6 dx = |u|66 ≤ C kuk6V ,

Ω

where the constant C > 0 is determined by the imbedding H10 (Ω ) ,→ L6 (Ω ). In the same way, we also see that, if u and v ∈ V, the products u2 , uv and v2 are in L3 (Ω ). Thus, if u and v are in a ball of V of radius R, by Minkowski’s inequality (see theorem A.57) we obtain ku3 − v3 k2 =

Z

|u − v|2 |u2 + uv + v2 |2 dx

Ω

≤

Z

|u − v|6 dx

1/3 Z

|u2 + uv + v2 |3 dx

2/3

Ω

Ω

= |u − v|26 |u2 + uv + v2 |23 2

≤ C ku − vk2V kuk4V + kvk4V

≤ C(R) ku − vkV .

This shows the asserted local Lipschitz continuity. We shall use proposition 3.15 also in the next section, for the proof of the uniqueness of solutions to problem (Hε ). We can now prove the asserted backward uniqueness result: PROPOSITION 3.16 Let T > 0, and assume that problem (P) has two solutions u, v, with u, v ∈ C([0, T ]; V) ∩ L2 (0, T ; H2 (Ω )) , ut , vt ∈ L2 (0, T ; H) , and u(T, ·) = v(T, ·). Then, u ≡ v on [0, T ]. PROOF We follow Temam, [Tem88, ch. 3.6]. Consider the difference z = u − v, which solves equation (3.46). Arguing by contradiction, assume that there is t0 ∈ [0, T [ such that z(t0 , ·) 6= 0. Then, by continuity, there is a largest interval [t0 ,t1 ] ⊂ [0, T [ such that z(t, ·) 6= 0 if t0 ≤ t < t1 ,

z(t1 , ·) = 0 .

Let M := maxt0 ≤t ≤t1 kz(t, ·)k. On the interval [t0 ,t1 [, the function t 7→ ln

M kz(t, ·)k

110

3

Attractors for Semilinear Evolution Equations

is well defined, nonnegative and differentiable. Using equation (3.46), we compute that (omitting the variables; recall convention 3.10) 1 M d d 2hz, zt i =− hz, z − (u3 − v3 ) + ∆zi ln = (− 1 ln kzk2 ) = − 21 dt kzk dt 2 kzk2 kzk2 = −1 +

k∇zk2 1 + hu3 − v3 , zi . kzk2 kzk2

(3.56)

Since u, v ∈ C([0, T ]; V), by proposition 3.15 we can estimate hu3 − v3 , zi ≤ Ck∇zk kzk ,

(3.57)

with C depending on u and v. Setting Λ (t) :=

k∇z(t, ·)k2 , kz(t, ·)k2

(3.58)

we deduce from (3.56) and (3.57) that M d ln ≤ −1 + Λ (t) +C(Λ (t))1/2 ≤ 41 C2 + 2Λ (t) . dt kz(t, ·)k

(3.59)

Integrating (3.59) in [t0 ,t], t0 < t < t1 , we obtain 0 ≤ ln

M M ≤ ln + 1 C(t − t0 ) + 2 kz(t, ·)k kz(t0 , ·)k 4

Z t

Λ (s) ds =: ϕ(t) .

(3.60)

t0

We shall show that ϕ remains bounded as t → t1− ; as a consequence, (3.60) yields the desired contradiction, since kz(t, ·)k → 0 as t → t1− . To prove that ϕ is bounded, it is sufficient to show that Λ satisfies the differential inequality dΛ ≤ CΛ dt

(3.61)

for almost all t ∈ [t0 ,t1 [, with C > 0 independent of t. Indeed, (3.61) implies that Λ (t) ≤ Λ (t0 )eC(t −t0 ) , so that

Z t t0

Λ (s) ds ≤ Λ (t0 ) C1 eC(t1 −t0 ) − 1 .

To show (3.61), we first observe that (omitting again the variables, as in convention 3.10) h−∆z − Λ z, zi = h−∆z − Λ z,Λ zi = 0

(3.62)

3.4

The Hyperbolic Problem

111

(recall that Λ is independent of x). Next, recalling (3.46), we compute that 1 1 dΛ h∇z, ∇zt ikzk2 − k∇zk2 hz, zt i = 2 dt kzk4 1 = (h−∆z, zt i − Λ hz, zt i) kzk2 1 h−∆z − Λ z, z − (u3 − v3 ) + ∆zi . = kzk2 By (3.62) and (3.57), we can proceed with 1 dΛ 1 h−∆z − Λ z, ∆z + Λ z − (u3 − v3 )i = 2 dt kzk2 1 −k∆z + Λ zk2 + h∆z + Λ z, u3 − v3 i = 2 kzk 1 ≤ − 21 k∆z + Λ zk2 + 21 ku3 − v3 k2 2 kzk 1 ≤ Ck∇zk2 , 2kzk2 from which (3.61) follows. This concludes the proof of the backward uniqueness for solutions of (3.1).

3.4 The Hyperbolic Problem In this section we consider the hyperbolic IBVP (Hε ). As mentioned in (3.13), the underlying space for the semiflow S generated by (3.4) is now the product space X := V × H, and each orbit is the image of a pair (u(·, u0 , u1 ), ut (·, u0 , u1 )). We also establish regularity results in the subspace X1 := D × V. We will consider in X various norms, having different weights depending on the parameter ε. All these norms will be equivalent to the graph norm in X = V × H, defined by k(u, v)k2X := kuk2V + εkvk2H = k∇uk2 + εkvk2 ,

(3.63)

which is itself weighted with respect to ε. For this norm we have a result analogous to proposition 3.5: PROPOSITION 3.17 Let V and H be as above, and ε > 0. Assume that the function u ∈ L2 (0, T ; V) is such that utt ∈ L2 (0, T ; V 0 ) and εutt − ∆u ∈ L2 (0, T ; H). Then the function R(t) := εkut (t)k2 + k∇u(t)k2

112

3

Attractors for Semilinear Evolution Equations

is absolutely continuous on [0, T ], and the identity d εkut (t)k2 + k∇u(t)k2 = 2hεutt (t) − ∆u(t), ut (t)i dt

(3.64)

holds in D0 (0, T ) and, in fact, for almost all t ∈ [0, T ]. A proof of this result when ε = 1 can be found in Temam, [Tem88, lem. II.4.1]; the case of variable ε follows in the same way. Note that the absolute continuity of R follows from (3.64), and the fact that both R and the function t 7→ 2hεutt (t) − ∆u(t), ut (t)i are in L1 (0, T ).

3.4.1 Step 1: The Solution Operator In this section we prove a global existence and well posedness result for problem (Hε ), which allows us to define the associated solution operator for any ε > 0. If the equation is autonomous, i.e. if f only depends on x, this solution operator is in fact a continuous semiflow. For simplicity, we consider in detail only the case when 0 < ε ≤ 1. DEFINITION 3.18 Let T > 0, u0 ∈ V, u1 ∈ H, f ∈ L2 (0, T ; H). A function u is a WEAK SOLUTION of problem (Hε ), with g(u) = u3 − u, on the interval [0, T ] if u ∈ C([0, T ]; V) ∩ C1 ([0, T ]; H), u(0, ·) = u0 , and for all ϕ ∈ W˜ 0 (T ) (the space defined in (3.49)), Z T

−εhut , ϕt i + hut , ϕi + h∇u, ∇ϕi + hu3 − u, ϕi dt

0

Z T

= 0

h f , ϕi dt + εhu1 , ϕ(0)i .

(3.65)

Note that the left side of (3.65) makes sense, since u3 ∈ L2 (0, T ; H) if u belongs to C([0, T ]; V). In fact, we now have Z T 0

ku3 (t)k2 dt ≤

Z T 0

|u(t)|66 dt ≤

Z T 0

ku(t)k6V dt ≤ kuk4C([0,T ];V )

Z T 0

ku(t)k2V dt .

We consider in X an ε-weighted norm whose square is defined by E0 (u, v) := εkvk2 + εhu, vi + 12 kuk2 + k∇uk2 .

(3.66)

This is indeed an equivalent norm, since by Schwarz’ inequality 2 1 2 k(u, v)kX

≤ E0 (u, v) ≤ αk(u, v)k2X ,

(3.67)

where α := max

n

3 1 2 , λ1

o +1 ,

(3.68)

3.4

The Hyperbolic Problem

113

with λ1 as in (3.16). In analogy to proposition 3.17, we have PROPOSITION 3.19 Let V and H be as above, and 0 < ε ≤ 1. Assume that the function u ∈ L2 (0, T ; V) is such that utt ∈ L2 (0, T ; V 0 ), and εutt − ∆u ∈ L2 (0, T ; H). Then the function t 7→ E0 (u(t), ut (t)) is absolutely continuous on [0, T ], and satisfies the identity d E0 (u(t), ut (t)) = hεutt (t) − ∆u(t), 2ut (t) + u(t)i dt + εkut (t)k2 + hut (t), u(t)i − k∇u(t)k2 in D0 (0, T ) and, in fact, almost everywhere in [0, T ]. We have then the following result: THEOREM 3.20 For all ε ∈ ]0, 1], f ∈ L2 (0, T ; H), u0 ∈ V and u1 ∈ H, there exists a unique weak solution u of problem (Hε ). u depends continuously on the initial data u0 , u1 , and the function t 7→ E0 (u(t), ut (t)) is absolutely continuous on [0, T ]. If in addition ft ∈ L2 (0, T ; H), u0 ∈ D and u1 ∈ V, then u ∈ C([0, T ]; D) ∩ C1 ([0, T ]; V) ∩ C2 ([0, T ]; H) , and the function t 7→ k∆u(t)k2 + kut (t)k2 + kεutt (t)k2 is absolutely continuous on [0, T ]. PROOF A proof of the existence and regularity part can again be given by means of a standard Galerkin approximation method, as in Lions, [Lio69], or Temam, [Tem88, ch. IV]. Since the procedure is similar to the one we followed for the parabolic problem (P), we do not repeat the details. In particular, the absolute continuity of the map t 7→ E0 (u(t), ut (t)) is a consequence of proposition 3.19. It is worth noting explicitly that in order to obtain uniform estimates on the norm E0 , the standard practice of multiplying the equation of (3.4) by 2ut in L2 (Ω ) is supplemented by a second multiplication of the equation by λ u, where λ is chosen conveniently. In our case, we choose either λ = 1 or λ = ε1 , according to the type of estimate that we need to establish. Moreover, we shall make constant use of the Sobolev imbeddings and interpolation inequalities reported in theorems A.74 and A.61. To prove the uniqueness of the solutions of (3.4), and the continuity of each operator S(t), we consider two solutions of (3.4), whose difference z satisfies the equation εztt + zt − ∆z = g(v) − g(u) ,

(3.69)

114

3

Attractors for Semilinear Evolution Equations

in the distributional sense of definition (3.18). Since the function u 7→ u3 is monotone, by propositions 3.19 and 3.15 we obtain (dropping the argument t and using different constants C) d E0 (z, zt ) = −(2 − ε)kzt k2 − k∇zk2 − hg(u) − g(v), 2zt + zi dt ≤ 2kg(u) − g(v)k kzt k + kzk2 ≤ Ck∇zk2 + kzt k2 + (1 +C)kzk2 ≤ C(k∇zk2 + kzk2 ) ≤ C1 E0 (z, zt ) , where the constants C and C1 depend on u and v. Applying proposition 2.63, we obtain E0 (z(t), zt (t)) ≤ E0 (z(0), zt (0))eC1 t ,

(3.70)

which implies the uniqueness and well-posedness in the large of solutions of (3.4). Since T is arbitrary in theorem 3.20, we obtain that if f ∈ L2loc (0, +∞; H), then u ∈ C([0, +∞[; V) ∩ C1 ([0, +∞[; H) . Thus, theorem 3.20 defines a unique global solution of problem (Hε ). If f is independent of t, theorem 3.20 shows that problem (Hε ) generates a continuous semiflow S on the space X = V × H, defined by (3.13). In particular, each operator S(t) is Lipschitz continuous, since from (3.70) we deduce that, for each fixed t > 0, E0 (S(t)(u0 , u1 ) − S(t)(v0 , v1 )) ≤ E0 ((u0 , u1 ) − (v0 , v1 ))eC1 t .

(3.71)

We also remark that the change of variable t → −t transforms the equation of (3.4) into εutt − ut − ∆u + g(u) = f˜ , with f˜(t, x) := f (−t, x). It is then easy to see that theorem 3.20 still holds, provided that f is defined in [−T, 0]. Thus, if f ∈ Cb (R; H), the solution of (3.4) is defined for all time. If in particular f (t) ≡ f ∈ H is independent of time, S is a flow. This is in sharp contrast to the parabolic problem, where S was only a semiflow. In fact, each operator S(t) is an isomorphism of X into itself, and also of X1 into itself. (This last assertion can be proved by means of further a priori estimates on the solution of (3.4); see e.g. Temam, [Tem88, ch. IV], for details.)

3.4.2 Step 2: Absorbing Sets As in section 3.3.2, we now proceed to establish time-independent a priori estimates on the solution u of (3.4). In order to take care of the nonlinear term u3 , for (u, v) ∈ X we introduce the function N0 (u, v) := E0 (u, v) + 12 |u|44 = εkvk2 + εhu, vi + 12 kuk2 + k∇uk2 + 12 |u|44 .

(3.72)

3.4

The Hyperbolic Problem

115

Because of (3.67), N0 is positive definite. We claim: PROPOSITION 3.21 Assume f ∈ Cb ([0, +∞[; H). There exists R0 > 0 such that the set B0 := {(u, v) ∈ X : N0 (u, v) ≤ R20 } is bounded, absorbing and positively invariant for the solution operator S in X = V × H. (Note that B0 is not a ball of X .) PROOF Multiplying the equation of (3.4) in H by 2ut + u (this is justified by proposition 3.19), we obtain (dropping the argument t) d εkut k2 + εhut , ui + 12 kuk2 + k∇uk2 + 12 |u|44 dt + (2 − ε)kut k2 + k∇uk2 + |u|44 = h f + u, 2ut + ui.

(3.73)

Setting F := supt ≥0 k f (t)k2 , we estimate h f , 2ut + ui ≤ 4F + 14 kut k2 + F + 41 kuk2 ≤ C1 + 14 kut k2 + 14 |u|44 ; hu, 2ut + ui ≤ 14 kuk2 + 14 kut k2 + kuk2 ≤ C2 + 14 |u|44 + 41 kut k2 . Thus, from (3.73) and (3.67) we obtain that d N0 (u, ut ) + 21 kut k2 + k∇uk2 + |u|44 ≤ C3 . dt Recalling then (3.67), and that ε ≤ 1 < α, we deduce the linear differential inequality d 1 N0 (u, ut ) ≤ C . N0 (u, ut ) + 2α dt Since the function t 7→ N0 (u(t), ut (t)) is absolutely continuous, by proposition 2.64 we deduce that for all t ≥ 0, N0 (u(t), ut (t)) ≤ (N0 (u0 , u1 ) − 2αC)e−t/2α + 2αC .

(3.74)

This implies not only the boundedness of the function t 7→ (u(t), ut (t)) on [0, +∞[, but also that, if R20 > 2αC, each corresponding set B0 is as claimed. In particular, if N0 (u0 , u1 ) > 2αC, we have that N0 (u(t), ut (t)) ∈ B0 for all t ≥ T0 , where T0 = 0 if N0 (u0 , u1 ) ≤ R20 , and T0 := 2α ln

N0 (u0 , u1 ) − 2αC R20 − 2αC

116

3

Attractors for Semilinear Evolution Equations

otherwise. This ends the proof of proposition 3.21. To conclude, we mention that in section 4.4.1 of chapter 4 we will show that if ε is sufficiently small, S also admits a positively invariant set B1 , bounded and absorbing in X1 . On the other hand, if ε > 1 we can still establish the existence of an absorbing set for S, with the same proof; however, instead of (3.66), we have to consider a differently weighted norm, whose square is defined by 1 (u, v) 7→ εkvk2 + hu, vi + 2ε kuk2 + k∇uk2

(otherwise, E0 is not necessarily positive definite), and the a priori estimates of proposition 3.21 are obtained by multiplying the equation of (3.4) by 2ut and ε1 u.

3.4.3 Step 3: Compactness of the Solution Operator Since in the hyperbolic case the solution operator S does not enjoy any smoothing property for t > 0 (on the contrary, S is an isomorphism in X ), we proceed instead to establish a decomposition of S like in proposition 2.48, under the condition that ε be sufficiently small. That is, we construct families S1 and S2 , with S1 uniformly compact for large t, S2 uniformly decaying to 0, and S = S1 + S2 . This construction is presented in Babin-Vishik, [BV92, sct. II.6], for the case when ε = 1 and g is a more general nonlinearity, satisfying conditions (3.9) and (3.10). Here, the smallness of ε allows us to considerably simplify their proofs. The family S2 = (S2 (t))t ≥0 of operators on X = V ×H is defined as follows. Given (u0 , u1 ) ∈ X and t ≥ 0, S2 (t)(u0 , u1 ) is the pair of functions in X defined by Ω 3 x 7→ (S2 (t)(u0 , u1 ))(x) := (v(t, x, u0 , u1 ), vt (t, x, u0 , u1 )) , where v(·, ·, u0 , u1 ) is the solution of the IBVP  3   εvtt + vt − ∆v + v = 0 v(0, ·) = u0 , vt (0, ·) = u1   v|∂ Ω = 0 .

(3.75)

Given this function v, we define a second function w(·, ·, u0 , u1 ) as the solution of the IBVP  3   εwtt + wt − ∆w = f + v − g(v + w) (3.76) w(0, ·) = 0 , wt (0, ·) = 0   w|∂ Ω = 0 . Since w depends on v, which in turn depends on (u0 , u1 ), problem (3.76) defines a family S1 = (S1 (t))t ≥0 of operators on X , by Ω 3 x 7→ (S1 (t)(u0 , u1 ))(x) := (w(t, x, u0 , u1 ), wt (t, x, u0 , u1 )) ,

3.4

The Hyperbolic Problem

117

for (u0 , u1 ) ∈ X and (t, x) ∈ [0, +∞[ × Ω . If v and w solve problems (3.75) and (3.76), we deduce that the function u := w + v solves problem (Hε ). Hence, the decomposition S = S1 + S2 does hold. We proceed then to show that the operators S1 and S2 so defined on X satisfy the requirements of the decomposition (2.38). Note that S2 is a semiflow, but S1 is not, because problem (3.76) is not autonomous. We now prove the uniform decay of S2 on bounded sets of X . PROPOSITION 3.22 Let (u0 , u1 ) ∈ V × H. Problem (3.75) has, for all ε ∈ ]0, 1], a unique solution v ∈ C([0, +∞[; V) ∩ C1 ([0, +∞[; H) . If α is defined as in (3.68), v satisfies the estimate N0 (v(t), vt (t)) ≤ N0 (u0 , u1 )e−t/2α .

(3.77)

PROOF Only (3.77) needs to be proven. The procedure is identical to the one we followed to obtain (3.74). Multiplying the equation in (3.75) by 2vt and v, and adding the resulting identities, we now obtain d 1 N0 (v, vt ) ≤ 0 , N0 (v, vt ) + 2α dt from which (3.77) follows. The uniform compactness of S1 for large t, at least when ε is sufficiently small, is a consequence of the second part of the following proposition. PROPOSITION 3.23 Let (u0 , u1 ) ∈ V × H, and let v be the function provided by proposition 3.22. Problem (3.76) has, for all ε ∈ ]0, 1], a unique solution w ∈ C([0, +∞[; V) ∩ C1 ([0, +∞[; H). If in addition ft ∈ Cb ([0, +∞[; H), there exists ε0 ∈ ]0, 1] such that for all ε ∈ ]0, ε0 ] and all t ≥ 0, w(t) ∈ H5/4 (Ω ) , wt (t) ∈ H1/4 (Ω ) .

(3.78)

Moreover, there is M > 0, independent of w and ε, such that for all t ≥ 0, εkwt (t)k2H1/4 (Ω ) + kw(t)k2H5/4 (Ω ) ≤ M 2 .

(3.79)

PROOF Only (3.78) and (3.79) need to be proved. In the estimates that follow, in accord with our convention 3.11, we denote by C, C1 , . . ., various generic positive constants which are independent of t and of the right side of (3.79). In particular, we recall that the norms defined by E0 (v, vt ) and E0 (w, wt ) are bounded with respect to

118

3

Attractors for Semilinear Evolution Equations

t, because of propositions 3.21 and 3.22. More precisely, consider the absorbing set B0 for S whose existence is assured by proposition 3.21. Recalling (3.74) and (3.77), and of course that w = u − v, by (3.67) we deduce that there exists C1 , depending on the norm (3.63) of (u0 , u1 ) in X , such that for all t ≥ 0 E0 (v(t), vt (t)) + E0 (w(t), wt (t)) ≤ C12 .

(3.80)

Note that, because of (3.67), and ε ≤ 1, C1 can be chosen independent of ε. We now show that w, solution of (3.76), is in fact more regular than v, namely that (3.78) holds for each t ≥ 0. This is essentially due to the fact that the initial values in (3.76) are (evidently!) more regular than (u0 , u1 ), and that the highest order term v3 does not appear in the equation of (3.76). To achieve this, we establish an additional estimate on a higher norm for w, which we define by resorting to the fractional powers of the operator L introduced in (5) of section 3.2.1. More precisely, recalling that L is self-adjoint and positive, we can define its fractional powers on H by means of its spectral basis (see section 3.2, as well as section A.5.5). In particular, we have the imbeddings dom(Li/8 ) ,→ Hi/4 (Ω ) ,→ L12/(6−i) (Ω ) ,

(3.81)

the first of which follows from theorem A.78, and the second, which is valid for 0 ≤ i < 6 (recall that n ≤ 3), from theorem A.69. We consider then in X1 = D × V a norm which has a different weight with respect to ε: more precisely, for (u, v) ∈ X1 we define 1 Eε (u, v) := εkL1/8 vk2 + hL1/8 v, L1/8 ui + 2ε kL1/8 uk2 + kL5/8 uk2 .

Note that Eε is indeed the square of an equivalent norm in H5/4 (Ω ) × H1/4 (Ω ). In fact, from (A.30) and (A.71) we immediately obtain that λ1 kL1/8 uk2 ≤ kL5/8 uk2 ; consequently, we easily deduce that, for all (u, v) ∈ X1 , 1/8 2 5/8 2 1/8 2 5/8 2 α 1 ≤ E (u, v) ≤ , εkL vk + kL uk εkL vk + kL uk ε 2 ε

(3.82)

(3.83)

where α is as in (3.68). Our goal is to show that the function t 7→ Eε (w(t), wt (t)) is bounded, uniformly in t and ε. To this end, we multiply the equation in (3.76) by 2L1/4 wt + ε1 L1/4 w. Setting Φ := f + v3 − g(v + w) ,

(3.84)

we obtain (omitting the variable t) d Eε (w, wt ) + kL1/8 wt k2 + ε1 kL5/8 wk2 = hΦ, 2L1/4 wt + ε1 L1/4 wi . dt

(3.85)

3.4

The Hyperbolic Problem

119

We rewrite (3.85) as d Eε (w, wt ) − 2hΦ, L1/4 wi + kL1/8 wt k2 + ε1 kL5/8 wk2 − α1 hΦ, L1/4 wi dt = −2hΦt , L1/4 wi + ε1 − α1 hΦ, L1/4 wi =: Λ . (3.86) To estimate Λ , we show that there exist positive constants C2 and C3 , depending on f and C1 of (3.80), but not on ε, such that 2hΦ, L1/4 wi ≤ C2 , 2hΦt , L

1/4

wi ≤

(3.87)

5/8 1 3 wk2 + 21 kL1/8 wt k2 . 2 C3 + 2ε kL

(3.88)

Indeed, since Φ = f − 3v2 w − 3vw2 − w3 + v + w ,

(3.89)

at first we have that 2hΦ, L1/4 wi ≤ C(k f k + kv2 w + vw2 + w3 k + kv + wk)kL1/4 wk ≤ C(k f k + |v|26 |w|6 + |v|6 |w|26 + |w|36 + kv + wk)k∇wk .

(3.90)

From this, (3.87) follows, recalling (3.80). Next, we compute from (3.89) that Φt = ft − (6vw + 3w2 − 1)vt − (3v2 + 6vw + 3w2 − 1)wt .

(3.91)

Therefore, setting F1 := maxt ≥0 k ft (t)k, we can estimate 2hΦt , L1/4 wi ≤ 2F1 kL1/4 wk +C |v|6 |w|12 + |w|28 + 1 |vt |2 |L1/4 w|4 +C |v|26 + |v|6 |w|6 + |w|26 + 1 |wt |12/5 |L1/4 w|4 .

(3.92)

We now resort to the interpolation inequality 1/2

1/2

|w|8 ≤ C|w|6 |w|12

(see theorem A.61), as well as to the imbedding inequalities (3.81) with i = 3, i = s and i = 1, which yield |L1/4 w|4 ≤ CkL1/4 wkH3/4 (Ω ) ≤ CkL1/4 wkdom(L3/8 ) ≤ CkL5/8 wk , |w|12 ≤ CkwkH5/4 (Ω ) ≤ CkL5/8 wk , |wt |12/5 ≤ Ckwt kH1/4 (Ω ) ≤ CkL1/8 wt k . Thus, recalling also (3.82), we can proceed from (3.92) with 1 2hΦt , L1/4 wi ≤ 2F +Ckv k +Ckw k kL5/8 wk t t λ 1

+C (k∇vk + k∇wk) kvt k kL5/8 wk2

120

3

Attractors for Semilinear Evolution Equations +C (k∇vk + k∇wk) k∇wk kL5/8 wk kL1/8 wt k =: R .

(3.93)

Recalling (3.80), we can estimate R ≤ C √1ε kL5/8 wk +C √1ε kL5/8 wk2 +CkL5/8 wk kwt k 1 1 kL5/8 wk2 +C √1ε kL5/8 wk2 + 6ε kL5/8 wk2 + 32 C2 εkL1/8 wt k2 . (3.94) ≤ 23 C2 + 6ε

At this point, we define 1 ε0 := min{ 36C 2 , 1} .

Then, if ε ≤ ε0 (this is our smallness assumption on ε), we deduce from (3.94) that 1 R ≤ 32 C3 + 2ε kL5/8 wk2 + 12 kL1/8 wt k2 ;

inserting this into (3.93), we obtain (3.88), with C3 = 32 C2 (independent of ε). Inserting (3.87) and (3.88) into (3.86), we obtain d 1 kL5/8 wk2 − α1 hΦ, L1/4 wi ≤ C4 ; Eε (w, wt ) − 2hΦ, L1/4 wi + 12 kL1/8 wt k2 + 2ε dt recalling (3.83) we deduce the linear differential inequality d 1 Eε (w, wt ) − 2hΦ, L1/4 wi + 2α Eε (w, wt ) − 2hΦ, L1/4 wi ≤ C4 . dt

(3.95)

Because of the initial conditions of (3.76), by proposition 2.63 we obtain that for all t ≥ 0, Eε (w(t), wt (t)) − 2hΦ(t), L1/4 w(t)i ≤ 2αC4 ; therefore, by (3.87), Eε (w(t), wt (t)) ≤ 2αC4 +C2 . Recalling (3.83) and (3.67) we conclude that, for all t ≥ 0, εkL1/8 wt (t)k2 + kL5/8 w(t)k2 ≤ 2(2αC4 +C2 ) =: C5 , with C5 independent of t and ε. This means that the orbit {(w(t), wt (t)) : t ≥ 0} = {S1 (t)(u0 , u1 ) : t ≥ 0} lies in a bounded set of 5/4

dom(L5/8 ) × dom(L1/8 ) = H0 (Ω ) × H1/4 (Ω ) ; thus, (3.78) and (3.79) follow, and the proof of proposition 3.23 is complete. Since H5/4 (Ω ) × H1/4 (Ω ) is compactly imbedded in X , we conclude that the family S1 is uniformly compact for large t.

3.4

121

The Hyperbolic Problem

3.4.4 Step 4: Conclusion If problem (Hε ) is autonomous, we can now deduce the existence of the global attractor for S, as a consequence of theorem 2.50. In conclusion, we have: THEOREM 3.24 Let u0 ∈ H10 (Ω ), and u1 , f ∈ L2 (Ω ). If ε is sufficiently small, the initial boundary value problem (Hε ) defines a semiflow S in X = H10 (Ω ) × L2 (Ω ). S admits a global attractor in X , given by the set A = Aε = ω(B0 ) =

\[

S(t)B0 ,

s≥0 t ≥s

where B0 is the absorbing set described in proposition 3.21. As in the parabolic case, the global attractor Aε so obtained contains, in addition to all stationary solutions of problem (Hε ), that is, the solutions of the elliptic problem (3.55), also the unstable manifolds (in X ) of all these stationary solutions (see e.g. Babin-Vishik, [BV92, sct. III.3]). The existence of Aε can also be established by means of α-contractions, and Aε is contained and bounded in the “more regular” space X1 = D × V. We show these two results in the next sections. Finally, the fractal dimension of Aε is finite.

3.4.5 Attractors via α -Contractions In this section we show how to establish the existence of the global attractor, again if ε is sufficiently small, by means of theorem 2.56 (note that ω(B0 ) is not empty, since it contains the stationary solutions of (3.4)). By proposition 3.21 the semiflow S admits a bounded, positively invariant absorbing set B0 . Recalling proposition 2.59, to apply theorem 2.56 it is sufficient to find an appropriate pseudometric δ on X , and a number t∗ > 0, such that condition (2.47) holds, with T = S(t∗ ). This is the goal of our next argument, for which the positive invariance of B0 is essential. For fixed T > 0 we define in X × X the function Z δT ((u, v), (u, ¯ v)) ¯ := 0

T

2

kP1 (S(t)(u, v)) − P1 (S(t)(u, ¯ v))k ¯ dt

1/2 ,

(3.96)

where P1 is the projection from V × H onto V. We explicitly remark that, although P1 (S(t)(u, v)) and P1 (S(t)(u, ¯ v)) ¯ are in V, in (3.96) we consider the norm of their difference in H. We claim: PROPOSITION 3.25 For each T > 0, δT is a pseudometric on X , precompact on B0 with respect to the norm of X defined by E0 .

122

3

Attractors for Semilinear Evolution Equations

PROOF The finiteness of δT follows from the invariance of B0 , since if (u, v) ∈ B0 , then so is the whole arc {S(t)(u, v) : 0 ≤ t ≤ T }. In particular, this means that the set {u(t) : 0 ≤ t ≤ T } is bounded in V, and therefore in H. The other conditions of a pseudometric are immediately verified. To see that δT is precompact on B0 , we need to show that, given any sequence ((un , vn ))n∈N ⊂ B0 (thus, bounded), there is a subsequence, still denoted ((un , vn ))n∈N , which converges relative to δT . Because of (3.96), this amounts to show that the subsequence (un )n∈N converges in L2 (0, T ; H). Thus, let ((un , vn ))n∈N ⊂ B0 . Since B0 is positively invariant, the corresponding orbits {S(t)(un , vn ) : t ≥ 0} remain in B0 . Since B0 is bounded, (3.67) implies that, setting wn (t) := P1 (S(t)(un , vn )) , there is M > 0, independent of ε, such that for all t ≥ 0 and n ∈ N, k∇wn (t)k2 + εkwtn (t)k2 ≤ M 2 .

(3.97)

We now recall from theorem A.82 that, since the injection of V into H is compact, the injection of ˜ ) := {u ∈ L2 (0, T ; V) : ut ∈ L2 (0, T ; H)} W(T into L2 (0, T ; H) is also compact. Hence, since (3.97) implies that (un )n∈N is bounded ˜ ), there is a subsequence of (un )n∈N which converges in L2 (0, T ; H), as in W(T desired. Our final step is to estimate the difference of two solutions of (3.4) in a sharper way than (3.70), so as to be able to apply proposition 2.59. To this end, letting U0 := (u0 , u1 ) and U 0 := (u¯0 , u¯1 ), we claim: PROPOSITION 3.26 There are ε0 ∈ ]0, 1] and K > 0, such that for all ε ≤ ε0 and U0 , U 0 ∈ B0 , t ≥ 0, Eε (S(t)U0 − S(t)U 0 ) ≤ e−t/2α Eε (U0 −U0 ) + ε1 K 2 (δt (U0 ,U 0 ))2 ,

(3.98)

with δt defined in (3.96), and α in (3.68). PROOF Let S(t)U0 =: (u(t), ut (t)) and S(t)U 0 =: (u(t), ¯ u¯t (t)). Since U0 and U 0 are in B0 , which is positively invariant, S(t)U0 and S(t)U 0 remain in B0 for all t ≥ 0. Since B0 is also bounded, there is M > 0, independent of ε, such that for all t ≥ 0, k∇u(t)k + k∇u(t)k ¯ ≤M.

(3.99)

Let z(t) := u(t) − u(t). ¯ Multiplying (3.69) by 2zt + ε1 z, we obtain (omitting, as usual, the variable t) d Eε (z(t), zt (t)) + kzt (t)k2 + ε1 k∇z(t)k2 dt

3.5

123

Regularity

= g(u(t)) ¯ − g(u(t)), 2zt (t) + ε1 z(t) .

(3.100)

Since g is Lipschitz continuous from V into H on bounded sets of V (this is a consequence of proposition 3.15), we deduce from (3.99) that there is K > 0, independent of t and ε, such that 2kg(u(t)) ¯ − g(u(t))k kzt (t)k ≤ 2Kk∇z(t)k kzt (t)k ≤ 2K 2 k∇z(t)k2 + 21 kzt (t)k2 , 1 ¯ − g(u(t))k kz(t)k ε kg(u(t))

≤ ε1 Kk∇z(t)k kz(t)k ≤

2 2 1 1 2 4ε k∇z(t)k + ε K kz(t)k .

Consequently, we obtain from (3.100) d Eε (z(t), zt (t)) + kzt (t)k2 + ε1 k∇z(t)k2 dt 1 ≤ 2K 2 + 4ε k∇z(t)k2 + 21 kzt (t)k2 + ε1 K 2 kz(t)k2 . Setting then ε0 :=

1 , 8K 2

we deduce that if ε ≤ ε0 ,

d 1 k∇z(t)k2 ≤ ε1 K 2 kz(t)k2 . Eε (z(t), zt (t)) + 21 kzt (t)k2 + 2ε dt

(3.101)

We easily verify that Eε (z, zt ) ≤ α kzt k2 + ε1 k∇zk2 ; hence, we obtain from (3.101) the linear differential inequality d 1 Eε (z(t), zt (t)) ≤ ε1 K 2 kz(t)k2 . Eε (z(t), zt (t)) + 2α dt By proposition 2.63, we deduce that Eε (z(t), zt (t)) ≤ e−t/2α Eε (z(0), zt (0)) + ε1 K 2

Z t

e−(t −θ )/2α kz(θ )k2 dθ ,

0

from which (3.98) follows. We can finally conclude the proof of the existence of a global attractor, by means of theorem 2.56 and proposition 2.59. In fact, choosing t∗ > 0 such that q := e−t∗ /2α < 1, from (3.98) we see that the operator T = S(t∗ ) and the pseudometric δt∗ satisfy condition (2.47) of proposition 2.59. Hence, T is an α-contraction. The fact that ω(B0 ) is the desired attractor for S follows then from theorem 2.56.

3.5 Regularity In this section, we prove a result concerning the regularity of the global attractor of the hyperbolic problem (Hε ), assuming again that the parameter ε is sufficiently small.

124

3

Attractors for Semilinear Evolution Equations

THEOREM 3.27 Let ε0 be as in proposition 3.23. For ε ∈ ]0, ε0 ], let Aε be the global attractor of problem (Hε ), given by theorem 3.24. There is ε1 ∈ ]0, ε0 ] such that if ε ≤ ε1 , Aε is contained in a bounded set of X1 . This set is independent of ε. PROOF We follow an idea presented in Grasselli-Pata, [GP02]; for an alternative proof, see e.g. Temam, [Tem88, sct. IV.6]. We proceed in three steps. At first we show, with methods similar to those of section 3.4.3, that the attractor Aε is contained in the subspace X1/4 := H5/4 (Ω ) × H1/4 (Ω ), and bounded with respect to the norm defined by k(u, v)k2X1/4 := εkL1/8 vk2 + kL5/8 uk2 ,

(u, v) ∈ X1/4 .

Then, we improve the estimates of solutions in X1/4 , and show that if ε is sufficiently small, a technique analogous to that of section 3.4.3 allows us to conclude that, in fact, Aε ⊆ X1 , and is bounded with respect to the norm whose square is defined by E1 (u, v) := E0 (∇u, ∇v) ,

(u, v) ∈ X1 .

Finally, we establish the bound k∇u1 k2 + k∆u0 k2 ≤ M 2 ,

(3.102)

where M > 0 is independent of ε and (u0 , u1 ) ∈ Aε . 1. Let (u0 , u1 ) ∈ Aε . Since Aε = ω(B), by proposition 2.15 there are sequences ((ϕn , ψn ))n∈N ⊆ B and (tn )n∈N ⊆ ]0, +∞[ such that S(tn )(ϕn , ψn ) → (u0 , u1 ) in X and tn → +∞ as n → ∞. Decompose S(tn ) = S1 (tn ) + S2 (tn ) as in section 3.4.3. Since B is bounded in X , (3.77) implies that there is R > 0, independent of ε, such that for all n ∈ N, N0 (S2 (tn )(ϕn , ψn )) ≤ R e−tn /2α ,

(3.103)

with N0 defined in (3.72). Since the function f is independent of t, (3.79) holds, so the sequence (S1 (tn )(ϕn , ψn ))n∈N is bounded in X1/4 . Since X1/4 is compactly imbedded into X , there is a subsequence, which we still denote by (S1 (tn )(ϕn , ψn ))n∈N , converging to some element (u¯0 , u¯1 ) weakly in X1/4 , and strongly in X . Because of (3.103), S(tn )(ϕn , ψn ) → (u¯0 , u¯1 ) in X ; hence, (u0 , u1 ) = (u¯0 , u¯1 ) ∈ X1/4 . Since (u0 , u1 ) was arbitrary in Aε , we conclude that Aε ⊆ X1/4 . Moreover, since we can now say that S1 (tn )(ϕn , ψn ) → (u0 , u1 ) weakly in X1/4 , k(u0 , u1 )kX1/4 ≤ lim inf kS1 (tn )(ϕn , ψn )kX1/4 , n→∞

(3.104)

3.5

Regularity

125

and this shows that Aε is bounded in X1/4 . 2. Set now, as in section 3.4.3, (u(t), ut (t)) := S(t)(u0 , u1 ) , (v(t), vt (t)) := S2 (t)(u0 , u1 ) , (w(t), wt (t)) := S1 (t)(u0 , u1 ) . In the first part of this proof we have shown that (u, ut ) ∈ Cb (R; X1/4 ). Since (u0 , u1 ) ∈ X1/4 , we can deduce, as in the proof of proposition 3.23, that (v, vt ) ∈ Cb ([0, +∞[; X1/4 ). Indeed, replacing the function Φ of (3.84) by −v3 , the same estimates that in that proof led to the exponential inequality (3.95) now lead to the inequality d 1 (Eε (v, vt ) − 2hv3 , L1/4 vi) + 2α (Eε (v, vt ) − 2hv3 , L1/4 vi) ≤ C4 . dt From this, we obtain that for all t ≥ 0, Eε (v(t), vt (t)) ≤ 2hv3 (t), L1/4 v(t)i + max{Eε (u0 , u1 ) − 2hu30 , L1/4 u0 i, 2αC4 } . Recalling that (u0 , u1 ) is in a bounded set of X1/4 , we easily conclude that (v, vt ) ∈ Cb ([0, +∞[; X1/4 ) as claimed. It follows then that also (w, wt ) = (u − v, ut − vt ) ∈ Cb ([0, +∞[; X1/4 ) . 3. We can now bootstrap the argument of the first part of this proof. We first show that if ε is sufficiently small, then, as a consequence of the fact that (u0 , u1 ) is in a bounded set of X1/4 , (w, wt ) ∈ Cb ([0, +∞[; X1 ). To this end, we (formally) multiply the equation of (3.76) in H by −2∆wt − ∆w: dropping as usual the argument t, we obtain d (εk∇wt k2 + εh∇w, ∇wt i + 12 k∇wk2 + k∆wk2 + 2hΦ, ∆wi) dt + k∇wt k2 + k∆wk2 + β2 hΦ, ∆wi D E1 2 = 2Φt + β − 1 Φ, ∆w , 1

(3.105)

with Φ as in (3.84) and β1 := 2(1 + λ1 ). To estimate the right side of (3.105), we 1 first note that, as in (3.90), recalling (3.80), hΦ, ∆wi ≤ C k∆wk ≤ C + 16 k∆wk2 .

(3.106)

Next, recalling (3.91) (with ft ≡ 0), 2hΦt , ∆wi ≤ C(|v|12 |w|∞ + |w|12 |w|∞ + 1)|vt |12/5 k∆wk +C(|v|212 + |v|12 |w|12 + |w|212 + 1)|wt |3 k∆wk .

(3.107)

126

3

Attractors for Semilinear Evolution Equations

Recalling (3.81) and theorem A.61, we have the imbeddings H5/4 (Ω ) ,→ L12 (Ω ) , H1/4 (Ω ) ,→ L12/5 (Ω ) as well as the interpolation inequalities 1/2

|w|∞ ≤ Ck∆2 wk1/2 |w|6 +C|w|6 , |wt |3 ≤ Ck∇wt k1/2 kwt k1/2 +C kwt k . By ellipticity (see theorem A.77), we can further estimate 1/2

1/2

1/2

|w|∞ ≤ CkwkH2 (Ω ) |w|6 +C|w|6

1/2

1/2

≤ C(k∆wk1/2 + kwk1/2 )|w|6 +C|w|6 .

Thus, recalling that both (v, vt ), (w, wt ) ∈ Cb ([0, +∞[; X1/4 ), we obtain from (3.107) that, for δ and η > 0, 2hΦt , ∆wi ≤ C(1 + δ k∆wk)k∆wk +Ck∇wt k1/2 k∆wk ≤ C + δ k∆wk2 + ηk∇wt k2 + ηk∆wk2 ,

(3.108)

where C denotes various constants, depending on δ , η and, of course, on the uniform bounds on (v, vt ) and (w, wt ) in X1/4 , but not on ε. Choosing δ and η sufficiently small, we obtain from (3.106) and (3.108), inserted into (3.105), that d (εk∇wt k2 + εh∇w, ∇wt i + 12 k∇wk2 + k∆wk2 + 2hΦ, ∆wi) dt + 12 k∇wt k2 + 21 k∆wk2 + β2 hΦ, ∆wi ≤ C . 1

(3.109)

As in (3.83), we easily verify that if ε ≤ min{ε0 , 13 } =: ε1 , Ψ := εk∇wt k2 + εh∇w, ∇wt i + 21 k∇wk2 + k∆wk2 + 2hΦ, ∆wi ≤ β1 12 k∇wt k2 + k∆wk2 + β2 hΦ, ∆wi ; 1

hence, (3.109) yields the exponential inequality Ψ 0 + β1 Ψ ≤ C . 1

From this, we obtain that for all t ≥ 0, εk∇wt (t)k2 + εh∇w(t), ∇wt (t)i + 21 k∇w(t)k2 + k∆w(t)k2 ≤ 2hΦ(t), ∆w(t)i +Cβ1 ≤ 2kΦ(t)k2 + 21 k∆w(t)k2 +Cβ1 . Since the function t 7→ kΦ(·,t)k is uniformly bounded in t, by Schwarz’ inequality we conclude that there is M1 > 0 such that for all t ≥ 0 and ε ≤ ε1 , εk∇wt (·,t)k2 + k∆w(·,t)k2 ≤ M12 .

(3.110)

3.5

127

Regularity

This shows that (w, wt ) ∈ Cb ([0, +∞[; X1 ) as claimed. 4. We now conclude as in the first part of this proof. Given arbitrary (u0 , u1 ) ∈ Aε , we construct sequences ((ϕn , ψn ))n∈N ⊆ B and (tn )n∈N as before. Estimate (3.110) implies that a subsequence of (S1 (ϕn , ψn ))n∈N converges to a limit (u¯0 , u¯1 ) weakly in X1 , and strongly in X . Again, (3.103) allows us to conclude that (u0 , u1 ) = (u¯0 , u¯1 ) ∈ X1 . Finally, as in (3.104), we deduce from (3.110) that (u0 , u1 ) is in a bounded set of X1 . Since (u0 , u1 ) is arbitrary in Aε , this implies that Aε is bounded in X1 . 5. We now proceed to show that, in fact, Aε can be bounded in X1 independently of ε. Let (u0 , u1 ) ∈ Aε . By proposition 2.39, (u0 , u1 ) lies on a complete orbit (u(t), ut (t))t ∈R , contained in Aε ; without loss of generality, we can assume that (u0 , u1 ) = (u(0), ut (0)). Since all the constants appearing in the proof of the boundedness of Aε in X1/4 and X1 , including in particular the constant R of (3.103), the boundedness implied by (3.104), and M1 of (3.110), depend only on the norm of u in Cb (R; V), we deduce that the estimate εk∇u1 k2 + k∆u0 k2 ≤ C2

(3.111)

holds uniformly with respect to ε and (u0 , u1 ) ∈ Aε . This provides part of (3.102); to remove the dependence of the term with u1 on ε, we now prove that there exists a positive constant C3 , independent of ε, such that for all t ∈ R and ε ∈ ]0, ε1 ], εkutt (t)k2 + k∇ut (t)k2 ≤ C3 .

(3.112)

To this end, we first note that, since Aε is invariant, from (3.111) we deduce that for all t ∈ R, 2

εk∇ut (t)k2 + k∆u(t)k ≤ C2 .

(3.113)

As a preliminary step, we establish the estimate kut (t)k ≤ C ,

(3.114)

where, here and in the sequel, we denote by C a generic positive constant, independent of ε ∈ ]0, ε1 ] and t ∈ R. Multiplying equation (3.4) by 2ut we obtain, by (3.113), ε

d kut k2 + kut k2 ≤ k f + ∆u + u − u3 k2 dt

≤ C k f k2 + k∆uk2 + kuk2 + k∇uk6 ≤ C .

(3.115)

Integrating (3.115) on an arbitrary interval [t0 ,t] ⊂ R, and recalling (3.113) again, we obtain kut (t)k2 ≤ C ε1 e−(t −t0 )/ε + 1 , from which we deduce (3.114) by letting t0 → −∞.

128

3

Attractors for Semilinear Evolution Equations

We now differentiate equation (3.4) with respect to t, and multiply the resulting equation by 2utt + ut , to obtain d 2 2 εkutt k2 + εhutt , ut i + 12 kut k2 + k∇ut k2 + (2 − ε) kutt k + k∇ut k dt

= − (3u2 − 1)ut , 2utt + ut =: R1 . (3.116) Combining the imbedding and elliptic estimates from theorems A.74 and A.77, and recalling (3.80) and (3.114), we can estimate the right side of (3.116) by R1 ≤ 2 |3u2 − 1|∞ kut k (kutt + kut k) (3.117) ≤ C 1 + kuk22 (1 + kutt k) ≤ 12 kutt k2 +C . We now denote by Φ1 (ut , utt ) the term under differentiation in (3.116), and note that, as in (3.67), (3.118) Φ1 (ut , utt ) ≥ α1 εkutt k2 + k∇ut k2 . Replacing (3.117) into (3.116), and recalling (3.114), we obtain, as usual, the inequality d Φ1 (ut , utt ) + α1 Φ1 (ut , utt ) ≤ C . dt Integrating this inequality on an arbitrary interval [t0 ,t] ⊂ R, we obtain Φ1 (ut (t), utt (t)) ≤ e−(t −t0 )/α Φ1 (ut (t0 ), utt (t0 )) + αC .

(3.119)

From (3.4) for t = t0 we have εutt (t0 ) = f (t0 ) + u(t0 ) − (u(t0 ))3 + ∆u(t0 ) − ut (t0 ) ; therefore, because of (3.113) and (3.114), Φ1 (ut (t0 ), utt (t0 )) ≤ ε1 C . Replacing this into (3.119) we obtain then that Φ1 (ut (t), utt (t)) ≤ C ε1 e−(t −t0 )/α + 1 .

(3.120)

Letting t0 → −∞, and recalling (3.118), we can finally deduce (3.112) from (3.120). 6. We can now conclude the proof of theorem 3.27: indeed, (3.102) follows from (3.111), and (3.112) with t = 0. REMARK 3.28 Following the proof of theorem 3.27, we realize that the requirement that ε be small is essential only in order to show that the boundedness of Aε in X1 is uniform with respect to ε. In fact, it is possible to show that each attractor

3.5

129

Regularity

Aε is bounded in X1 . To this end, it is sufficient to modify the estimates we have established, multiplying the equation e.g. by 2ut + λε u, or by 2ut + λ u, with λ sufficiently small. Indeed, both Temam, [Tem88], and Grasselli-Pata, [GP02], prove the boundedness of Aε in X1 for ε = 1. On the other hand, if ε is small, we can actually prove much more than mere boundedness in X1 : THEOREM 3.29 In the same conditions of theorem 3.27, if ε ≤ ε1 each attractor Aε is compact in X1 . PROOF Let (ϕn , ψn )n∈N be a sequence in Aε . Since Aε is bounded in X1 , there is a subsequence, still denoted by (ϕn , ψn )n∈N , converging to a pair (ϕ∗ , ψ∗ ), weakly in X1 and strongly in X . We must show that (ϕn , ψn ) → (ϕ∗ , ψ∗ )

strongly in X1 .

1. By proposition 2.39, for each n ∈ N there is a complete orbit (un (t), utn (t))t ∈R passing through (ϕn , ψn ), and completely contained in Aε ; without loss of generality, we can assume that (un (0), utn (0)) = (ϕn , ψn ) .

(3.121)

From (3.112) and (3.113) we deduce that, for all T > 0, the functions un , utn and uttn are, respectively, in a bounded set of L2 (−T, T ; H2 (Ω )), L2 (−T, T ; H1 (Ω )), and L2 (−T, T ; L2 (Ω )) (this set depends on ε, but here we consider ε as a fixed parameter). Hence, by theorem A.16, for each ε there are a function u and a subsequence, still denoted by (un )n∈N , such that un → u utn → ut uttn → utt

in in in

L2 (−T, T ; H2 (Ω )) 2

1

2

2

L (−T, T ; H (Ω ))

weakly ,

(3.122)

weakly ,

(3.123)

L (−T, T ; L (Ω )) weakly .

(3.124)

We recall from theorem A.82 that for each j ∈ N, the injections {u ∈ L2 (−T, T ; H j+1 (Ω )) : ut ∈ L2 (−T, T ; H j (Ω ))} ,→ L2 (−T, T ; H j (Ω )) (3.125) are compact. Taking j = 1, (3.125), (3.122) and (3.123) imply that un → u

in

L2 (−T, T ; H1 (Ω )) strongly .

(3.126)

Likewise, taking j = 0 in (3.125), (3.123) and (3.124) imply that utn → ut

in L2 (−T, T ; L2 (Ω ))

strongly .

(3.127)

130

3

Attractors for Semilinear Evolution Equations

We now show that we can deduce, from (3.126), that (un )3 → u3

L2 (−T, T ; L2 (Ω ))

in

strongly .

(3.128)

In fact, we can estimate Z T

k(un )3 − u3 k2 dt ≤

Z T −T

−T

|(un )2 + un u + u2 |2∞ kun − uk2 dt ;

(3.129)

since un is bounded in L∞ (−T, T ; H2 (Ω )), and this space is continuously imbedded into L∞ (Ω × ] − T, T [), (3.128) follows from (3.129). 2. Recall now the definition (3.18) of weak solution of (3.4): in particular, un solves (3.65). Then, by (3.126), (3.127) and (3.128), letting n → +∞ we deduce that u is also a solution of (3.65). We now show that (u(t), ut (t)) = Sε (t)(ϕ∗ , ψ∗ ) ,

t ∈ R.

(3.130)

To this end, we recall from theorem A.81 that the space {u ∈ L2 (−T, T ; H2 (Ω )) : ut ∈ L2 (−T, T ; L2 (Ω ))} is continuously injected in C([−T, T ]; H1 (Ω )), with max kun (t) − u(t)k1

−T ≤t ≤T

1/2

1/2

≤ C kun − ukL2 (−T,T ;H2 (Ω )) kutn − ut kL2 (−T,T ;L2 (Ω )) + kun − ukL2 (−T,T ;L2 (Ω )) . Hence, (3.122) and (3.127) imply that un → u

in

C([−T, T ]; H1 (Ω )) strongly .

(3.131)

From theorem A.81, we also have the trace estimate max kutn (t) − ut (t)k

(3.132)

−T ≤t ≤T

1/4

3/4

≤ C kun − ukL2 (−T,T ;H1 (Ω )) kuttn − utt kL2 (−T,T ;L2 (Ω )) + kun − ukL2 (−T,T ;L2 (Ω )) . Hence, by (3.131), we conclude that also utn → ut

in C([−T, T ]; L2 (Ω ))

strongly .

(3.133)

From (3.131) and (3.133) we deduce that u(0) = ϕ∗ and ut (0) = ψ∗ ; hence, (3.130) follows, by the uniqueness of solutions to (Hε ) with the same initial values. 3. Consider now the difference zn := un − u, which solves the equation εzttn + ztn − ∆zn = zn − ((un )3 − u3 ) .

(3.134)

Multiplying (3.134) in L2 (Ω ) by −∆(2ztn + zn ), we obtain d E0 (∇zn , ∇ztn ) + (2 − ε)k∇ztn k2 + k∆zn k2 = (un )3 − u3 − zn , −2∆(ztn + zn ) dt

3.5

Regularity

= 2 ∇((un )3 − u3 − zn ), ∇ztn − (un )3 − u3 − zn , ∆zn =: R2 ,

131 (3.135)

where E0 is as in (3.66). We can estimate the right side of (3.135) by R2 ≤ (2 + λ1 )k∇((un )3 − u3 − zn )k2 + 21 k∇ztn k2 + 12 k∆zn k2 ; thus, recalling (3.67), we obtain from (3.135) that d E0 (∇zn , ∇ztn ) + α1 E0 (∇zn , ∇ztn ) ≤ (2 + λ1 )k∇((un )3 − u3 − zn )k2 . dt

(3.136)

We now write ∇((un )3 − u3 − zn ) = ∇ ((un )2 + un u + u2 − 1)zn , and recall that, since n ≤ 3, H2 (Ω ) is an algebra (see theorem A.69). Hence, (un )2 + un u + u2 is in a bounded set of H2 (Ω ), and we easily see that k∇((un )3 − u3 − zn )k ≤ C k∇zn k , for suitable constant C depending only on Aε . Thus, we further obtain from (3.136) that d E0 (∇zn , ∇ztn ) + α1 E0 (∇zn , ∇ztn ) ≤ C (2 + λ1 )k∇zn k2 . dt

(3.137)

From (3.131) we deduce that for all T > 0 and all η > 0 there is n0 ∈ N such that for all n ≥ n0 and t ∈ [−T, T ], αC(2 + λ1 )k∇zn (·,t)k2 ≤ η . Thus, integrating (3.137) in [−T, T ], for n ≥ n0 , we obtain that E0 (∇zn (t), ∇ztn (t)) ≤ e−T /α E0 (∇zn (−T ), ∇ztn (−T )) + η ≤ C1 e−T /α + η ,

(3.138)

with C1 again depending only on Aε , because (zn , ztn ) is in a bounded set of X1 . Given then η > 0, we can therefore predetermine T > 0 so large that C1 e−T /α ≤ η; with this choice of T , we determine n0 as above, and (3.138) implies that for all n ≥ n0 and t ∈ [−T, T ], E0 (∇zn (t), ∇ztn (t)) ≤ 2η .

(3.139)

Recalling (3.67), and choosing t = 0 in (3.139), we deduce that un (0) → u(0) in H2 (Ω ) , utn (0) → ut (0) in H1 (Ω ) . Since (un (0), utn (0)) = (ϕn , ψn ) and (u(0), ut (0)) = (ϕ∗ , ψ∗ ), we conclude that (ϕn , ψn ) converges strongly in X1 to (ϕ∗ , ψ∗ ), as desired. This concludes the proof of theorem 3.29.

132

3

Attractors for Semilinear Evolution Equations

3.6 Upper Semicontinuity of the Global Attractors 1. The hyperbolic problem (Hε ) can be seen as a perturbation, for small ε > 0, of the reduced parabolic problem (P). In this context, a natural question is that of the convergence of the solutions of (Hε ), which we now rename uε to emphasize their dependence on ε, to the solution u of (P). Another, related question is whether the corresponding global attractors Aε and A, obtained in theorems 3.24 and 3.13, can be compared, and in what sense. In this section, we briefly consider some aspects of the latter question, proving the so-called UPPER SEMICONTINUITY of the global attractors as ε → 0. As a byproduct, we also obtain some results on the convergence uε → u, limited to motions contained in the attractors Aε and A. We note that, in general, the convergence uε → u can be expected to be singular in time, because of the loss of the initial condition on ut ; we refer to Lions, [Lio73], for an extensive presentation of results on singular perturbation problems of this type. In contrast, we shall see that if we restrict our attention to motions in the attractors, it is possible to transform the singular convergence problem into a regular convergence one. On the other hand, the question of the convergence of the global attractors Aε to A as ε → 0 is also complicated by the obvious fact that Aε is in the product space H10 (Ω ) × L2 (Ω ), while A is contained in H10 (Ω ) only. One way to solve this difficulty is to consider the projections P1 Aε of Aε into H1 (Ω ), and to show that lim ∂ (P1 Aε , A) = 0 ,

ε →0

where ∂ is the semidistance in H1 (Ω ) defined in (2.2) (see e.g. the survey article [EM95]). Another possibility, which is the one we present here, is to construct a “natural” imbedding of A into a set A0 ⊆ X = H10 (Ω ) × L2 (Ω ), and then show that lim ∂ (Aε , A0 ) = 0 ,

ε →0

where ∂ is now the semidistance in X . In this section, we loosely follow the presentation of Hale, [Hal88, sct. 4.10]; for more results on this type of question, and related ones, such as for instance the LOWER SEMICONTINUITY of the attractors, we refer to Hale-Raugel, [HR88, HR90], or to Grasselli-Pata, [GP02]. 2.

We recall the definition of upper and lower semicontinuity of a family of sets.

DEFINITION 3.30 Let X be a complete metric space, Λ ⊆ R, and (C λ )λ ∈Λ a family of subsets of X . Let λ0 ∈ Λ . Then: 1. (C λ )λ ∈Λ is UPPER SEMICONTINUOUS at λ0 if lim ∂ (C λ , C λ0 ) = 0 ,

λ →λ0

(3.140)

3.6

Upper Semicontinuity of the Global Attractors

133

2. (C λ )λ ∈Λ is LOWER SEMICONTINUOUS at λ0 if lim ∂ (C λ0 , C λ ) = 0 .

λ →λ0

3. (C λ )λ ∈Λ is CONTINUOUS at λ0 if it is upper and lower semicontinuous at λ0 . 3. We proceed then to prove the upper semicontinuity of the attractors Aε as ε → 0. Let A be the attractor of the semiflow S generated by the parabolic problem (P). As we remarked after theorem 3.13, A is bounded in H2 (Ω ); hence, we can introduce the set A0 := {(u, v) ∈ X : u ∈ A, v = f + u − u3 + ∆u} ,

(3.141)

which we consider as a “natural” imbedding of A in X . We have then the following result: THEOREM 3.31 Let ε1 be as in theorem 3.27 and, for 0 < ε ≤ ε1 , let Aε be the global attractor of the semiflow Sε generated by the hyperbolic problem (Hε ). Let A0 be as in (3.141). The family (Aε )0≤ε ≤ε1 is upper semicontinuous at ε = 0, with respect to the topology of X. PROOF Recalling (3.140), we must show that sup

inf

0

k∇(u − u)k ¯ 2 + kv − vk ¯ 2

1/2

→0

(3.142)

¯ v) ¯ ∈A (u,v)∈Aε (u,

as ε → 0. We reason by contradiction. Assuming (3.142) did not hold, we could find η0 > 0, and sequences (εn )n∈N ⊂]0, ε1 ], ((ϕn , ψn ))n∈N ⊆ Aεn , such that εn → 0, and for all n ∈ N, (3.143) k∇(ϕn − u)k ¯ 2 + kψn − vk ¯ 2 ≥ η02 . inf (u, ¯ v) ¯ ∈A0

By (3.102), we have the uniform estimate kψn k21 + kϕn k22 ≤ M 2 , with M independent of n; thus, there is a subsequence, still denoted by ((ϕn , ψn ))n∈N , converging to a limit (ϕ∗ , ψ∗ ) weakly in X1 and, by compactness, strongly in X . We now claim that (ϕ∗ , ψ∗ ) ∈ A0 : if true, this would contradict (3.143). We now proceed in analogy to the proof of theorem 3.29. By proposition 2.39, for each n ∈ N there is a complete orbit (un (t), utn (t))t ∈R = (Sεn (t)(ϕn , ψn ))t ∈R

(3.144)

134

3

Attractors for Semilinear Evolution Equations

contained in Aεn and passing through (ϕn , ψn ). In particular, we can assume that (3.121) holds, for un defined as in (3.144). From (3.112) and (3.113) we have the uniform estimates εn kuttn (t)k2 + kutn (t)k21 + kun (t)k22 ≤ M22 ,

(3.145)

with M2 independent of t and εn . From this it follows that, for all T > 0, the func√ tions uεn , utεn and εn uttεn are, respectively, in a bounded set of L∞ (−T, T ; H2 (Ω )), L∞ (−T, T ; H1 (Ω )) and L∞ (−T, T ; L2 (Ω )). Consequently, there is a function u, and a subsequence, still denoted (εn )n∈N , such that uεn → u

in L2 (−T, T ; H2 (Ω )) weakly ∗ ,

utεn → ut

in L2 (−T, T ; H1 (Ω )) weakly ∗ ,

εn uttεn → 0

in L2 (−T, T ; L2 (Ω )) weakly ∗ .

Proceeding exactly as in theorem 3.29, we show that u is a weak solution of the parabolic problem (P) on R. By the analogous of (3.131), ϕn = un (0) → u(0) in H1 (Ω ); hence, u(0) = ϕ∗ , and, therefore, u(0) ∈ H2 (Ω ). Moreover, since u is a complete orbit of S through ϕ∗ , by proposition 2.39 ϕ∗ ∈ A. By (3.145), √ √ √ kεn uttn (0)k = εn k εn uttn (0)k ≤ εn M2 ; hence, εn uttn (0) → 0 in L2 (Ω ). Consequently, utn (0) = f + un (0) − (un (0))3 + ∆un (0) − εn uttn (0) = f + ϕn − ϕn3 + ∆ϕn − εn uttn (0) → f + ϕ∗ − ϕ∗3 + ∆ϕ∗

(3.146)

in L2 (Ω ) weakly. Since utn (0) = ψn , from (3.146) we deduce that ψ∗ = f + ϕ∗ − ϕ∗3 + ∆ϕ∗ .

(3.147)

Since ϕ∗ ∈ A, (3.147) implies that (ϕ∗ , ψ∗ ) ∈ A0 , as claimed. Having thus reached a contradiction with (3.143), the proof of theorem 3.31 is complete. Note that, comparing (3.146) with (3.147), we have that utεn (0) → ut (0) . This means that, as mentioned above, the convergence uεn → u is not singular at t = 0.

Chapter 4 Exponential Attractors

In this chapter we give the definition of EXPONENTIAL ATTRACTOR for a semiflow S defined on a Hilbert space X , and describe an explicit construction of the exponential attractor when the semiflow satisfies a geometric property, called the DISCRETE SQUEEZING PROPERTY . As an application, we show that the continuous semiflows defined by the initial-boundary value problems (P) and (Hε ) (the latter at least for small ε) satisfy the discrete squeezing property. Therefore, these systems admit an exponential attractor.

4.1 Introduction 1. In the following, X is, as usual, a Banach space on R, with norm k · k. DEFINITION 4.1 Let G be an arbitrary subset of X , and T be one of the sets R, R≥0 , N or Z. A subset E ⊂ G is an EXPONENTIAL ATTRACTOR for the semiflow S = (S(t))t ∈T in G, if E is a compact, positively invariant set which has finite fractal dimension and attracts each bounded set B ⊆ G exponentially. The last requirement means that there are positive numbers c and k, depending on B, such that for all t ∈ T≥0 , ∂ (S(t)B, E) ≤ ce−kt ,

(4.1)

where ∂ is the semidistance in X introduced in definition 2.1. 2. A large class of dynamical systems is known to admit an exponential attractor; for example, the semiflows generated by the 2-dimensional Navier-Stokes equations with periodic boundary conditions, the Kuramoto-Sivashinski equations, the ChafeeInfante and the Cahn-Hilliard equations (in any space dimension), and the “original” Burger’s equation all admit an exponential attractor; we will see some examples in chapter 6. In contrast, much fewer systems are known to have an inertial manifold; for example, even for the 2-dimensional Navier-Stokes equations the existence of an inertial manifold is open.

135

136

4

Exponential Attractors

3. Exponential attractors share with attractors the property of being attracting sets, but they do not have to be invariant. However, in contrast to (2.33), which carried no information on the rate of convergence of the orbits to the attractor, (4.1) gives an explicit information on the rate of convergence of the orbits to the exponential attractor. The fact that this rate is exponential makes exponential attractors much more robust than global attractors for numerical analysis. Consequently, exponential attractors are generally more suitable for the qualitative study of the long-time behavior of a semiflow. For example, it is proven in Eden, Foias, Nicolaenko and Temam, [EFNT94, ch. 4], that, in the framework of Galerkin approximations schemes, approximate and exact exponential attractors are continuous with respect to the Hausdorff distance of definition 2.1. 4. When the global attractor A and an exponential attractor E exist, then A is contained in E. Indeed, since A is bounded, it is attracted by E. Recalling (2.33), this means that lim ∂ (S(t)A, E) = 0 .

t →+∞

(4.2)

Since A is invariant, (4.2) implies that ∂ (A, E) = 0. Thus, since E is compact, proposition 2.2 implies that A⊆E. 5. One of the assumptions under which we will prove the existence of an exponential attractor is that the semiflow S admits a compact, positively invariant absorbing set B. In this case, S also admits the global attractor A = ω(B). Then, as we have seen, A ⊆ E and, therefore, A has finite fractal dimension, with dimF (A) ≤ dimF (E) by (2.49). Thus, A can be imbedded in RN and it can be proven that N ≥ 2 dimF (A) + 1. Moreover, there exists a N-dimensional system of ODEs x˙ = F(x)

(4.3)

(with F not necessarily continuous) such that this imbedding of A is the global attractor (respectively, this imbedding of E is the exponential attractor) of the generalized semiflow (“generalized” because F may not be continuous) generated by (4.3) (see [EFNT94, thms. 10.2, 10.3]). We refer to [EFNT94] for more details. 6. Exponential attractors can in fact be seen, in some sense, as intermediate objects between attractors and inertial manifolds, which we study in detail in chapter 5. Essentially, inertial manifolds are positively invariant submanifolds of X , which also attract orbits exponentially. The basic difference between inertial manifolds and exponential attractors is that exponential attractors need not have a smooth structure, and inertial manifolds need not be compact. Because exponential attractors are not required to have a smooth structure, we can establish existence results for exponential attractors under conditions on the semiflow that are much less restrictive than those required for the existence of an inertial manifold. In fact, as shown in [EFNT94], not

4.2

The Discrete Squeezing Property

137

only there are semiflows, generated by PDEEs, for which the existence of an exponential attractor is known, while that of an inertial manifold is not, but there actually are semiflows that admit an exponential attractor, and not an inertial manifold. Indeed, in chapter 7 we present an example due to Mora and Solà-Morales, [MSM87], of a semiflow generated by an IBVP for a hyperbolic equation like (3.4), which does not admit any C1 inertial manifold. Of course, there are much simpler examples of two-dimensional ODEs which do not have an inertial manifold; for example, the system x˙ = x , y˙ = 2y , whose orbits admit two unstable manifolds, but no inertial manifold.

4.2 The Discrete Squeezing Property In this section we give the definition of the DISCRETE SQUEEZING PROPERTY, which is the main ingredient for the construction of an exponential attractor. The discrete squeezing property for the semiflow S generated by an evolution equation of the form ut + Au = f (u)

(4.4)

is usually defined by means of orthogonal projections in X . Roughly speaking, the discrete squeezing property translates a dichotomy principle, whereby either the system is exponentially contracting on a fixed compact set B ⊆ X , or the evolution of the difference of two solutions originating in B, when expressed as a Fourier series with respect to the eigenvectors of the operator A in (4.4), can be controlled by a finite number of terms of the series. In other words, in this series the tail can be dominated by its complementary finite sum. In more suggestive terms, this is often expressed by saying that if the system is not already exponentially contracting from the outset, then the higher modes of the difference of any two solutions can be dominated by the lower modes.

4.2.1 Orthogonal Projections As usual, we denote by k · k, h· , ·i and I, respectively the norm, the scalar product and the identity in X . Recall that a PROJECTION in X is a linear map P : X → X such that P2 = P. Suppose there exists a closed subspace W ⊂ X such that X can be decomposed into the direct sum of W and its orthogonal complement W ⊥ , i.e. X = W ⊕ W⊥ . Then, every x ∈ X can be expressed in a unique way as a sum x = w + z , w ∈ W , z ∈ W⊥ ,

(4.5)

138

4

Exponential Attractors

with hw, zi = 0 (by orthogonality). Consequently, kxk2 = kwk2 + kzk2 .

(4.6)

We define then the ORTHOGONAL PROJECTION P : X → W by P(x) = w ,

(4.7)

with w as in (4.5). This makes sense because w is unique. P is indeed a projection, since for all x ∈ X , P2 (x) = P(w) = w = P(x) ; that is, P2 = P. Note that P is linear, and (4.6) can be written as kxk2 = kP(x)k2 + k(I − P)(x)k2 .

(4.8)

From (4.8) we deduce that orthogonal projections are Lipschitz continuous: indeed, kP(x) − P(y)k2 = kP(x − y)k2 = kx − yk2 − k(I − P)(x − y)k2 ≤ kx − yk2 , and analogously for I − P. In fact, orthogonal projections are contractions, albeit not strict ones. Finally, we recall that orthogonal projections allow us to define a special cone in X , namely CP := {x ∈ X : k(I − P)(x)k ≤ kP(x)k} , or, equivalently (this is an immediate consequence of (4.8)) √ CP = {x ∈ X : kxk ≤ 2kP(x)k} .

(4.9)

(4.10)

4.2.2 Squeezing Properties To define the squeezing properties we want to introduce, we consider projections (4.7) defined into a subspace W ⊆ X which is finite dimensional. In this case, the opening of the cone CP is controlled by a projection of finite rank. We start with the definition of squeezing property for a map: DEFINITION 4.2 A map S : X → X satisfies the SQUEEZING PROPERTY relative to a subset B ⊆ X if there is a finite dimensional closed subspace XN ⊂ X and a corresponding orthogonal projection PN : X → XN , such that the following condition holds: There exists γ ∈ ]0, 21 [ such that for any u, v ∈ B, either kS(u) − S(v)k ≤ γku − vk ,

(4.11)

or, otherwise, S(u) − S(v) ∈ CPN , i.e. k(I − PN )(S(u) − S(v))k ≤ kPN (S(u) − S(v))k .

(4.12)

4.2

The Discrete Squeezing Property

139

This condition means that either S is contractive on the set B, or that any difference S(u) − S(v) is in the cone CPN . Note that, because of (4.9) [or (4.10)], the condition S(u) − S(v) ∈ / CPN can be expressed by either one of the conditions kPN (S(u) − S(v))k < k(I − PN )(S(u) − S(v))k , √ kS(u) − S(v)k > 2kPN (S(u) − S(v))k .

(4.13) (4.14)

Thus, the discrete squeezing property can also be expressed by the requirement that if u, v ∈ B are such that S(u) − S(v) ∈ / CPN , i.e. if either (4.13) or (4.14) holds, then S is contractive on B, i.e. (4.11) holds. We can then extend the definition of squeezing property to a semiflow: DEFINITION 4.3 Let B ⊆ X be bounded, and S = (S(t))t ∈T a semiflow in X . S satisfies the DISCRETE SQUEEZING PROPERTY relative to B if there is t∗ ∈ T , with t∗ > 0, such that the map S∗ := S(t∗ ) satisfies the squeezing property of definition 4.2. We will see that, in contrast to the case of inertial manifolds, the discrete squeezing property, i.e. the “cone condition” (4.12), is a dichotomy at the specific time t∗ only. That is, this condition neither requires nor implies that the squeezing property of definition 4.2 be satisfied by any of the other maps S(t), either at the intermediate times 0 < t < t∗ , or at all later times t > t∗ (that is, the invariance of the cone (4.9) for t > t∗ ). This explains in part the qualification of “discrete” in definition 4.3.

4.2.3 Squeezing Properties and Exponential Attractors The importance of the discrete squeezing property resides in the fact that it provides the main sufficient condition for the existence of an exponential attractor. The basic result concerns discrete semiflows generated by iterated maps: THEOREM 4.4 Assume the map S : X → X is Lipschitz continuous, and B ⊆ X is a nonempty, compact, and positively invariant subset. Assume that S satisfies the squeezing property of definition 4.2, relative to B. Then the discrete semiflow S˜ := (Sn )n∈N admits an exponential attractor E in B. Moreover, given any θ ∈ ]2γ, 1[, the fractal dimension of E (see (2.48)) can be estimated by √

n ln( 2 2L + 1) o θ −2γ dimF (E) ≤ N max 1, , − ln θ

(4.15)

where N is the rank of the projection appearing in definition 4.2, and L is the Lipschitz constant of S on B. We postpone the proof of theorem 4.4 to the end of this chapter. By means of theorem 4.4, we can construct an exponential attractor for a general semiflow:

140

4

Exponential Attractors

THEOREM 4.5 Assume the continuous semiflow S = (S(t))t ∈T satisfies the discrete squeezing property, relative to a nonempty, compact, positively invariant subset B ⊆ X . Let t∗ be as in definition 4.3, set S∗ := S(t∗ ), and let E∗ be the exponential attractor in X for the semiflow generated by S∗ , as in theorem 4.4. Assume further that S satisfies the following properties: PS1) For all t ∈ [0,t∗ ], S(t) is Lipschitz continuous from B into B, with Lipschitz constant L(t), L : [0,t∗ ] → ]0, +∞[ being a bounded function. PS2) For all x ∈ E∗ , the map S(·)x is Lipschitz continuous from [0,t∗ ] into B, with Lipschitz constant L0 (x), L0 : E∗ → ]0, +∞[ being also a bounded function. Then, the set E :=

[

S(t)E∗

(4.16)

0≤t ≤t∗

is an exponential attractor for S in B. If in addition B is an absorbing set which attracts all bounded sets of X , then E is an exponential attractor in X . Moreover, the fractal dimension of E (see (2.48)) can be estimated by √

n ln( 2 2L(t∗ ) + 1) o θ −2γ dimF (E) ≤ 1 + N max 1, , − ln θ

(4.17)

where N is the rank of the projection appearing in definition 4.2, L(t∗ ) is as in (PS1), and θ ∈ ]2γ, 1[. From this theorem, which we prove in the next section, it follows that one of the most direct ways to establish the existence of an exponential attractor for a given semiflow S is to show the existence of a compact, positively invariant set B, relative to which the system satisfies the discrete squeezing property. In most cases, B is a bounded absorbing set, whose existence is implied by the dissipativity of the system. The compactness of B is usually proven by showing that B ⊆ X1 , X1 a compactly imbedded subspace of X . We will follow this procedure for both problems (P) and (Hε ); however, we mention that, as shown in Eden, Foias and Kalantarov, [EFK98], the assumption of compactness of B is not actually needed, since in both cases S is an α-contraction, as we saw in chapter 3. This possibility of relaxing the requirement of compactness of B is of course extremely important for applications to situations when, as for hyperbolic equations, the solution operator is not regularizing. At any rate, once a compact, positively invariant set B has been determined, we proceed to conveniently choose a time t∗ , such that the map S(t∗ ) satisfies the squeezing property relative to B. The evolution of the system on [0,t∗ ] is then controlled by continuity, since the system is well posed on compact intervals (this “explains” (4.16)), while for t > t∗ the system is essentially finite dimensional. Finally, we remark that condition (PS1) was already assumed in theorem 2.56 (with E∗ replaced by A∗ ), and is usually a consequence of estimates that show that the Cauchy problem

4.2

141

The Discrete Squeezing Property

relative to the differential equation that generates the semiflow is well posed. On the other hand, condition (PS2) is a consequence of the fact that, usually, the compactness of B results from its being bounded in a subspace X1 compactly imbedded in X . In particular, since the solutions we consider are valued in B, this implies that these solutions are differentiable in t, with derivative still valued in X ; therefore, they are locally Lipschitz continuous in t as well.

4.2.4 Proof of Theorem 4.5 Recalling definition 4.1, we need to show that the set E defined by (4.16) is compact, positively invariant, satisfies (4.1), and has finite fractal dimension. 1. The compactness of E is proven exactly as in part 1 of the proof of theorem 2.56. 2. To show that E is positively invariant, we first note that for all t ≥ 0, [

S(t)E ⊆

S(θ )E∗ .

(4.18)

t ≤θ ≤t∗ +t

Indeed, let x ∈ E. There is then τ ∈ [0,t∗ ] such that x ∈ S(τ)E∗ . Thus, S(t)x ∈ S(t + τ)E∗ , and if θ = t + τ, S(t)x ∈ S(θ )E∗ , with t ≤ θ ≤ t∗ + t. This shows (4.18). Fix now t > 0, and consider first the case when 0 ≤ t ≤ t∗ . From (4.18) we obtain that ! ! S(t)E ⊆

[

[

S(θ )E∗ =

t ≤θ ≤t∗ +t

S(θ )E∗ ∪

t ≤θ ≤t∗

|

[

S(θ )E∗

.

t∗ ≤θ ≤t∗ +t

{z := E2

} |

{z := E3

}

Obviously, E2 ⊆ E. As for E3 , setting τ = θ − t∗ we can write E3 =

[

S(τ)S(t∗ )E∗ ;

0≤τ ≤t

and since E∗ is positively invariant with respect to S∗ = S(t∗ ), E3 ⊆

[

S(τ)E∗ ⊆ E .

0≤τ ≤t

Thus, S(t)E ⊆ E if 0 ≤ t ≤ t∗ . If instead t > t∗ , we can decompose t = nt∗ + θ , for some n ∈ N and θ ∈ [0,t∗ ]. Then, S(t)E = S(nt∗ )S(θ )E, and by the first part of this argument, S(θ )E ⊆ E. Consequently, recalling (4.18), S(t)E ⊆ S(nt∗ )E ⊆

[ nt∗ ≤s≤nt∗ +t∗

S(s)E∗ .

142

4

Exponential Attractors

Decomposing s = nt∗ +τ, with 0 ≤ τ ≤ t∗ , and recalling that E∗ is positively invariant with respect to S∗ = S(t∗ ), we obtain S(t)E ⊆

[

S(τ)S(nt∗ )E∗ =

0≤τ ≤t∗

[

[

S(τ)S∗n E∗ ⊆

0≤τ ≤t∗

S(τ)E∗ = E .

0≤τ ≤t∗

This completes the proof that E is positively invariant. 3. We now show that E attracts all bounded subsets of B exponentially. Since E∗ is an exponential attractor for S∗n in B, there are two positive constants c1 and k1 , depending on B, such that for all n ∈ N ∂ (S∗n B, E∗ ) ≤ c1 e−k1 n .

(4.19)

Let then G ⊆ B be bounded, and fix t ≥ t∗ . Given any x ∈ S(t)G, and z∗ ∈ E∗ , let g ∈ G be such that x = S(t)g, and decompose t = nt∗ + θ , for suitable n ∈ N, and θ ∈ [0,t∗ ]. Let z¯ := S(θ )z∗ . Then, z¯ ∈ E, and recalling (2.42), we can estimate kx − z¯k = kS(t)g − z¯k = kS(θ )S(nt∗ )g − S(θ )z∗ k ≤ L∗ kS∗n g − z∗ k .

(4.20)

In fact, recalling that B is positively invariant with respect to S∗n , and that G ⊆ B, we have that S(nt∗ )g = S∗n g ∈ S∗n (B) ⊆ B; moreover, also z∗ ∈ E∗ ⊆ B. From (4.20), it follows that inf kx − zk ≤ kx − z¯k ≤ L∗ kS∗n g − z∗ k . z∈E

Since z∗ is arbitrary in E∗ , inf kx − zk ≤ L∗ inf kS∗n g − z∗ k .

z∈E

z∗ ∈E∗

(4.21)

Since S∗n g ∈ B, recalling the definition of semidistance, we can proceed from (4.21) with inf kx − zk ≤ L∗ sup

z∈E

inf kb − z∗ k = L∗ ∂ (S∗n B, E∗ ) .

n B z∗ ∈E∗ b∈S∗

(4.22)

Since (4.22) is true for arbitrary x ∈ S(t)G, it follows that sup inf kx − zk = ∂ (S(t)G, E) ≤ L∗ ∂ (S∗n B, E∗ ) .

x∈S(t)G z∈E

(4.23)

Recalling (4.19), we deduce from (4.23) that ∂ (S(t)G, E) ≤ L∗ c1 e−k1 n .

(4.24)

Now, from t = nt∗ + θ ≤ nt∗ + t∗ , we deduce the inequality −n ≤ 1 − t/t∗ ; hence we deduce from (4.24) that ∂ (S(t)G, E) ≤ L∗ c1 e2k1 e−k1 t/t∗ . Thus, (4.1) holds, with c = L∗ c1 e2k1 and k = k1 /t∗ .

4.3

The Parabolic Problem

143

4. Finally, we prove estimate (4.17). To this end, note that E is the image of [0,t∗ ]× E∗ under the map F defined in step 1 of the proof of theorem 2.56. Since F is easily to seen to be Lipschitz continuous on [0,t∗ ]× E∗ as a consequence of the assumptions PS1) and PS2) on S), by (2.52) and (2.51) of proposition 2.61 we have dimF (E) ≤ dimF ([0,t∗ ] × E∗ ) ≤ dimF (E∗ ) + dimF ([0,t∗ ]) . Recalling (4.15), and that, by part 3 of proposition 2.61, the fractal dimension of an interval is 1, we conclude that (4.17) holds. This ends the proof of theorem 4.5.

4.3 The Parabolic Problem In this section and the next we proceed to show how theorem 4.5 can be applied to deduce the existence of an exponential attractor for the semiflows defined by the IBV problems (P) and (Hε ). Recall that both these systems have a bounded, positively invariant absorbing ball B, as well as a compact attractor A, as shown in chapter 3. With the same notations of chapter 3, we consider the parabolic IBVP (P) in X = L2 (Ω ) = H. In section 3.3 we proved that (P) defines a semiflow S on X , which admits a bounded, positively invariant absorbing set B. We now proceed to show that S admits an exponential attractor E ⊆ B. As outlined in the previous section, we proceed in two steps, first showing that B contains a compact, positively invariant set B1 , then showing that S satisfies the discrete squeezing property relative to B1 .

4.3.1 Step 1: Absorbing Sets in X1 We first show that B contains an absorbing set bounded in X1 = V = H10 (Ω ), and therefore compact in X . We claim: PROPOSITION 4.6 Assume f ∈ Cb ([0, +∞[; X ). The semiflow defined by the IBVP (P) admits a positively invariant absorbing set B1 , bounded in X1 . PROOF We must show that orbits originating in bounded sets of X1 enter, and eventually remain, in a set B1 as stated. To this end, we establish further a priori estimates on the solution of problem (P). Formally multiplying equation (3.1) in X by −2∆u (again, this procedure should be carried out at the level of Galerkin approximations), we obtain (omitting as usual the variable t) d k∇uk2 + 2k∆uk2 + 6hu2 ∇u, ∇ui = 2h f + u, −∆ui . dt

144

4

Exponential Attractors

Neglecting the positive term 6hu2 ∇u, ∇ui and integrating the last term by parts, we further obtain d k∇uk2 + 2k∆uk2 ≤ 2k f k k∆uk + 2k∇uk2 . dt From this we deduce that d k∇uk2 + k∆uk2 ≤ k f k2 + 2k∇uk2 . dt

(4.25)

We now recall estimate (3.51) of chapter 3, which we rewrite as d kuk2 + 2k∇uk2 + |u|44 ≤ k f k2 + 3kuk2 ≤ k f k2 +CΩ + |u|44 . dt Summing this to (4.25) we obtain d (kuk2 + k∇uk2 ) + k∆uk2 ≤ 2k f k2 +CΩ ≤ K , dt

(4.26)

where K := 2 supt ≥0 k f (·,t)k2 +CΩ is independent of t and u0 . From (4.26) we can deduce a linear differential inequality on the norm of u in X1 . Indeed, from the Poincaré inequalities (3.16) and (3.17) we have k∆uk2 ≥ β (kuk2 + k∇uk2 )

with β :=

λ12 . 1 + λ1

Inserting this into (4.26) we conclude that d (kuk2 + k∇uk2 ) + β (kuk2 + k∇uk2 ) ≤ K , dt which is the desired linear differential inequality. We can then conclude the proof of proposition 4.6 as in that of proposition 2.64.

4.3.2 Step 2: The Discrete Squeezing Property We now assume that f is independent of t, i.e. f (·,t) ≡ f ∈ H, and proceed to show that theorem 4.5 can be applied to the corresponding semiflow S generated by (3.1). THEOREM 4.7 Let B1 as in proposition 4.6. The semiflow defined by the IBVP (P) satisfies the discrete squeezing property relative to B1 . Consequently, the semiflow generated by problem (P) admits an exponential attractor in X . PROOF We consider the spaces XN and the corresponding orthogonal projections PN and QN defined in (3.26) and (3.27) of section 3.2.3, and refer to the alternative characterization of the discrete squeezing property “(4.13) =⇒ (4.11)”. In fact, we

4.3

The Parabolic Problem

145

prove a bit more: namely, that, given any t∗ > 0 and γ ∈ ]0, 12 [, there exists an integer N∗ , with the property that if u0 , v0 ∈ B1 are such that S(t∗ )u0 − S(t∗ )v0 ∈ / CN∗ , i.e. if kPN∗ (S(t∗ )u0 − S(t∗ )v0 )k < kQN∗ (S(t∗ )u0 − S(t∗ )v0 )k

(4.27)

(that is, if (4.13) holds for the operator S(t∗ )), then (4.11) must hold, i.e. kS(t∗ )u0 − S(t∗ )v0 k ≤ γku0 − v0 k.

(4.28)

Therefore, S satisfies the discrete squeezing property, relative to B1 . To this end, following Eden, Foias, Nicolaenko and Temam, [EFNT94, ch. 3], we study the evolution of the so-called “quotient norm” Λ (t) :=

k∇z(t)k2 , kz(t)k2

introduced in (3.58), where z(t) := S(t)u0 − S(t)v0 is the difference of the two solutions of (3.1) with initial data u0 and v0 . Our goal is to establish a series of inequalities for Λ , from which we can deduce (4.28). At first we note that, since the projections PN and QN are orthogonal in both X and X1 , if an inequality like (4.27) holds for generic N ∈ N and t ≥ 0, then, by (3.28), Λ (t) =

kPN (z(t))k21 + kQN (z(t))k21 kQN (z(t))k21 ≥ ≥ 21 λN+1 . 2 2 kPN (z(t))k + kQN (z(t))k 2kQN (z(t))k2

(4.29)

Next, we recall that Λ satisfies the differential inequality (3.61), i.e. Λ 0 ≤ CΛ , for almost all t ≥ 0. Integrating this inequality for 0 < s < t yields Λ (s) ≥ Λ (t)e−C(t −s) ; integrating a second time (with respect to s in the interval [0,t]) we obtain Z t 0

Λ (s) ds ≥ Λ (t)

Z t 0

eC(s−t) ds = C1 Λ (t)(1 − e−Ct ) .

(4.30)

Our last step is an estimate of z in terms of the integral at the left side of (4.30). Recalling (3.47), z satisfies, for almost all t > 0, the inequality d kzk2 + 2k∇zk2 = −2hg(u) − g(v), zi ≤ 2kzk2 . dt Thus, recalling the definition of Λ , we obtain from (4.31) d kz(t)k2 + 2Λ (t)kz(t)k2 ≤ 2kz(t)k2 ; dt

(4.31)

146

4

Exponential Attractors

applying proposition 2.63 we obtain, for t ≥ 0, Z t 2 2 kz(t)k ≤ kz(0)k exp 2t − 2 Λ (s) ds .

(4.32)

0

We are now ready to conclude. Replacing (4.30) into (4.32), we obtain that, if t > 0 is such that (4.29) holds, then kz(t)k2 ≤ kz(0)k2 exp 2t − C1 λN+1 (1 − e−Ct ) .

(4.33)

Recalling that λN → +∞ as n → +∞, given t∗ > 0 and γ ∈ ]0, 12 [ we choose N∗ so large that − C1 λN∗ +1 (1 − e−Ct∗ ) ≤ 2 ln γ − 2t∗

(4.34)

(note that the right side of (4.34) is negative). With this choice of N∗ , (4.33) implies that kz(·,t∗ )k ≤ γkz(·, 0)k that is, (4.28) holds. As we have already remarked, conditions (PS1) and (PS2) of theorem 4.5 hold, because problem (P) is well posed, and the corresponding semiflow is differentiable. Thus, the proof of theorem 4.7 is complete. REMARK 4.8 In estimate (4.31), we have made use of the fact that the nonlinearity g(u) = u3 − u satisfies hg(u) − g(v), u − vi ≥ −ku − vk2 . More generally, the right side of (4.31) is estimated by 2Lkzk2 , where L is the Lipschitz constant of g, and (4.33) becomes 1 λN+1 (1 − e−Ct ) ; kz(t)k2 ≤ kz(0)k2 exp 2Lt − 2C note that C also depends on L. N∗ is then chosen so that − 12 λN∗ +1 1 − e−Ct∗ ≤ 2 ln γ − 2Lt∗ . In any case, the choice of t∗ is independent of B1 . Actually, in the proof of theorem 4.7 we see that the set B1 can be replaced by the set B. In fact, the compact set B1 is needed only because it is required in theorem 4.5. In conclusion, the only relevant property we have actually needed is the Lipschitz continuity of the nonlinearity g, locally in V. It follows that the local Lipschitz continuity of g in V is a sufficient condition for the existence of an exponential attractor for the semiflow generated by a semilinear parabolic PDE of the form (3.1).

4.4

The Hyperbolic Problem

147

4.4 The Hyperbolic Problem We now turn to the hyperbolic IBV problem (Hε ) in X := H10 (Ω ) × L2 (Ω ). In section 3.4 we proved that (3.4) defines a semiflow S on X , which admits a bounded, positively invariant absorbing set B0 . We now proceed to show that S admits an exponential attractor E ⊆ B0 . Exactly as in the previous section, we proceed in two steps, first showing that B0 contains a compact, positively invariant set B1 , and then that S satisfies the discrete squeezing property relative to B1 . As in chapter 3, we set H := L2 (Ω ), V := H10 (Ω ), and choose in X the norm k(u, v)k2X := k∇uk2 + εkvk2 . We also consider the subspaces D := H2 (Ω ) ∩ H10 (Ω ) and X1 := D × V. Since we can choose on D the norm kukD = k∆uk, we consider in X1 the norm k(u, v)k2X1 := k∆uk2 + εk∇vk2 . This norm is equivalent to the norm defined by 1 E1 (u, v) := εk∇vk2 + h∇v, ∇ui + 2ε k∇uk2 + k∆uk2 .

(We recall that to distinguish between pairs of functions (u, v) and their scalar product in L2 (Ω ), the latter is denoted by hu, vi.)

4.4.1 Step 1: Absorbing Sets in X1 We first show that, at least if ε is sufficiently small, B0 contains an absorbing set bounded in X1 , and therefore compact in X . We claim: PROPOSITION 4.9 Assume f is independent of t, i.e. f (t) ≡ f ∈ H. Let λ1 be as in (3.16), and ε1 := min{1, λ1 }. For all ε ∈ ]0, ε1 ], the semiflow defined by the IBVP (3.4) admits a 1 positively invariant absorbing set B1 , bounded in X1 . PROOF We must show that orbits originating in bounded sets of X1 enter, and eventually remain, in a set B1 as stated. To this end, as in the parabolic case, we establish further a priori estimates on the solution of (3.4). Multiplying (formally) the equation of (3.4) in H by −2∆ut − ε1 ∆u, and adding the term 21 λ1 h f , ∆ui to both sides, we obtain d (E1 (u, ut ) + 2h f , ∆ui) + k∇ut k2 + ε1 k∆uk2 + 12 λ1 h f , ∆ui dt = 2hu3 − u, ∆ut i + hρ f + u3 − u, ε1 ∆ui ,

(4.35)

148

4

Exponential Attractors

with ρ := 12 ελ1 − 1. Recalling proposition 3.21, we have a bound on k∇u(·,t)k independent of t and ε. Consequently, we can estimate the right side of (4.35) as follows. At first, using the elliptic estimate ku(·,t)kH2 (Ω ) ≤ C(k∆u(·,t)k + ku(·,t)k) (see theorem A.77), and denoting by C various different positive constants, depending on the uniform bound on k∇u(·,t)k, but not on t nor on ε, we have 2hu3 − u, ∆ut i = −2h∇u − 3u2 ∇u, ∇ut i ≤ C(k∇uk + |u|26 |∇u|6 )k∇ut k ≤ C 1 + k∇uk2 kuk2 k∇ut k ≤ C(1 + k∆uk)k∇ut k 1 ≤ 21 k∇ut k2 + 4ε k∆uk2 + ε1 C

(4.36)

(recall that ε ≤ 1). Next, analogously: 1 ε hρ

f + u3 − u, ∆ui ≤ ε1 (kρ f k + |u|36 + kuk)k∆uk ≤ ε1 (kρ f k + k∇uk3 + kuk)k∆uk 1 k∆uk2 . ≤ C ε1 k∆uk ≤ C ε1 + 4ε

Replacing this and (4.36) into (4.35) we obtain d (E1 (u, ut ) + 2h f , ∆ui) + 12 (k∇ut k2 + ε1 k∆uk2 ) + 12 λ1 h f , ∆ui ≤ ε1 C . dt Again, we check that, since ε ≤

(4.37)

1 λ1 ,

E1 (u, v) ≤

2 2 1 2 λ1 (k∇vk + ε k∆uk )

for (u, v) ∈ X1 . Consequently, from (4.37) we derive the exponential inequality d (E1 (u, ut ) + 2h f , ∆ui) + 14 λ1 (E1 (u, ut ) + 2h f , ∆ui) ≤ ε1 C . dt From this we deduce, by proposition 2.63, that for all t ≥ 0 E1 (u(t), ut (t)) + 2h f , ∆u(t)i ≤ E1 (u0 , u1 ) + 2h f , ∆u0 i − λ4Cε e−λ1 t/4 + λ4Cε . 1

As usual, this implies that for all R >

4C λ1 ε ,

1

the set

B1 := {(u, v) ∈ X1 : E1 (u, v) + 2h f , ∆ui ≤ R} ∩ B is bounded, absorbing and positively invariant for S. This ends the proof of proposition 4.9.

4.4

The Hyperbolic Problem

149

4.4.2 Step 2: The Discrete Squeezing Property We now show that, again if ε is sufficiently small, theorem 4.5 can be applied to the semiflow S generated by (3.4). THEOREM 4.10 Let B1 and ε1 be as in proposition 4.9. There exists ε2 ∈]0, ε1 ] such that if ε ∈]0, ε2 ], the semiflow defined by the IBVP (3.4) satisfies the discrete squeezing property, relative to B0 . Consequently, the semiflow generated by problem (Hε ) admits an exponential attractor in X . PROOF We follow Eden, Milani and Nicolaenko, [EMN92]. Here too, we refer to the alternative characterization of the discrete squeezing property “(4.13) =⇒ (4.11)”. We consider the same spaces XN and projections PN , QN introduced in (3.26) and (3.27), and define corresponding product projections on X = V × H by PN : X → (PN (V) × PN (H)), QN = I − PN in the canonical way, i.e. PN (u, v) := (PN (u), PN (v)) , QN (u, v) = (QN (u), QN (v)) .

(4.38)

We must find t∗ and N∗ ∈ N such that if (u0 , u1 ), (v0 , v1 ) ∈ B1 are such that S(t∗ )(u0 , u1 ) − S(t∗ )(v0 , v1 ) ∈ / CN∗ , i.e. if kPN∗ (S(t∗ )(u0 , u1 ) − S(t∗ )(v0 , v1 ))kX < kQN∗ (S(t∗ )(u0 , u1 ) − S(t∗ )(v0 , v1 ))kX , (4.39) then (4.11) must hold for some γ ∈ ]0, 12 [, i.e. kS(t∗ )(u0 , u1 ) − S(t∗ )(v0 , v1 )kX ≤ γk(u0 , u1 ) − (v0 , v1 )kX .

(4.40)

In fact, we shall prove slightly more: namely, that for all t∗ > 0 and γ ∈ ]0, 21 [ there exist ε2 ∈ ]0, ε1 ] such that for all ε ∈ ]0, ε2 ], there exists N∗ ∈ N such that (4.40) holds. To this end, we define on X the function M(u, v) := hu, viH + k(u, v)k2X = εkvk2 + k∇uk2 + hu, viH , and show that M is the square of an equivalent norm in QN (X ). PROPOSITION 4.11 Assume ε ≤ 1, and let N ∈ N be such that ελN+1 ≥ 1 (this is possible since λN → +∞). Then for all (u, v) ∈ QN (X ), k(u, v)k2X ≤ 2 M(u, v) ≤ 3k(u, v)k2X .

(4.41)

150

4

Exponential Attractors

PROOF This result is a consequence of proposition 3.6 and Schwarz’ inequality. Indeed, if (u, v) ∈ QN (X ), by (3.28) hu, viH ≤

2 2 1 1 2ε kuk + 2 εkvk

≤

2 2 1 1 2ελN+1 k∇uk + 2 εkvk .

We now estimate the difference of two solutions of (3.4), whose orbits are in B0 . If u and v are two such solutions, corresponding to initial values U0 := (u0 , u1 ), V0 := (v0 , v1 ) ∈ B1 , we set z(t) := u(·,t) − v(·,t), and Z(t) := (z(t), zt (t)). At first, we recall estimate (3.71) of section 3.4.1, which provides a control of the growth of Z on bounded time intervals. More precisely, kZ(t)kX ≤ kZ(0)kX ect ,

(4.42)

for suitable c > 0 independent of the solutions u and v. Next, we establish a linear differential inequality on Z in QN (X ). PROPOSITION 4.12 Let N ∈ N be as in proposition 4.11, and ϕN := QN (z), ΦN := (ϕN , (ϕN )t ) = QN (Z). Then M(ΦN ) satisfies, for almost all t ≥ 0, the linear differential inequality d 1 M(ΦN (t)) + 2ε M(ΦN (t)) ≤ Kk∇z(t)k2 , dt

(4.43)

where K is independent of t and ε. PROOF The procedure is similar to the proof of proposition 3.21. Applying QN to the equation satisfied by z, i.e εztt + zt − ∆z = g(v) − g(u) , and noting that QN commutes with −∆, we see that, abbreviating ϕN = ϕ, ϕ satisfies the equation εϕtt + ϕt − ∆ϕ = QN (g(v) − g(u)) =: G .

(4.44)

Multiplying (4.44) in H by 2ϕt + ε1 ϕ and splitting one of the terms in two, we obtain d 1 1 M(Φ) + kϕt k2 + 2ε hϕt , ϕi + ε1 k∇ϕk2 = − 2ε hϕt , ϕi + hG, 2ϕt + ε1 ϕi . (4.45) dt {z } | {z } | =:H1

=:H2

By Schwarz’ inequality and proposition 3.6, we estimate −1 k∇ϕk2 H1 ≤ 41 kϕt k2 + 4ε12 kϕk2 ≤ 41 kϕt k2 + 4ε12 λN+1

4.4

The Hyperbolic Problem

1 ≤ 14 kϕt k2 + 4ε k∇ϕk2 ,

151 (4.46)

where we have used ελN+1 ≥ 1. To estimate H2 , we recall that, by proposition 3.21, u(·,t) and v(·,t) remain in a bounded set of V for all t ≥ 0, with bounds independent of ε if ε is small. Thus, recalling proposition 3.15, we can estimate 1 1 H2 ≤ 2kGk kϕt k + kϕk ≤ 2C k∇zk kϕt k + kϕk ε ε −1 ≤ C1 k∇zk2 + 14 kϕt k2 + 4ε12 λN+1 k∇ϕk2 ,

(4.47)

where C and C1 depend on the uniform bounds on k∇u(·,t)k and k∇v(·,t)k, but not on t nor on ε. Inserting (4.46) and (4.47) into (4.45), we easily deduce (4.43). We are now ready to conclude the proof of theorem 4.10. From (4.42) and (4.43) we obtain d 1 M(ΦN ) ≤ K E0 (z(t)) ≤ K E0 (z(0))ec t . M(ΦN ) + 2ε dt Integrating, we deduce that M(ΦN (t)) ≤ M(ΦN (0))e−t/2ε + 2εK E0 (z(0))ec t .

(4.48)

Thus, recalling (4.41), kΦN (t)k2X ≤ 3kΦN (0)k2X e−t/2ε + 6εK E0 (z(0))ec t ≤ 3E0 (z(0))(e−t/2ε + 2εK ec t ) .

(4.49)

Given then t∗ > 0 and γ ∈ ]0, 21 [, we first choose ε2 ∈ ]0, ε1 ] so small that 3(e−t∗ /ε2 + 2ε2 K ect∗ ) ≤ 12 γ 2 ; then, given ε ∈ ]0, ε2 ], we choose N∗ so large that λN∗ +1 ≥ ε1 .

(4.50)

With these choices, (4.49) implies that kΦN∗ (t∗ )k2X ≤ 12 γ 2 kZ(0)k2X .

(4.51)

Recalling that ΦN = QN (Z), from (4.39) and (4.51) we deduce that kZ(t∗ )k2X ≤ 2kΦN∗ (t∗ )k2X ≤ γ 2 kZ(0)k2X , that is, (4.40) holds. As a consequence of proposition 4.12, we can then conclude the proof of theorem 4.10, by means of theorem 4.5. Indeed, as we have remarked for the parabolic problem (P), assumptions (PS1) and (PS2) are satisfied, because problem (Hε ) is well posed in X , and the corresponding semiflow S is differentiable. Finally, note that, by (4.50), in general N∗ → +∞ if ε → 0.

152

4

Exponential Attractors

REMARK 4.13 1. As in the parabolic case (see remark 4.8), the proof of theorem 4.10 does not require the existence of the compact, positively invariant absorbing set B1 . This is only required by theorem 4.5, and is the only reason we have introduced proposition 4.9. However, this condition is not a necessary one, as shown in Eden, Foias and Kalantarov, [EFK98]. 2. As was the case for the global attractors, a natural question is to know whether any result on the upper or lower semicontinuity would hold for the exponential attractors of problems (Hε ) and (P), whose existence we have established in theorems 4.7 and 4.10. In the case of the global attractors, we were able to answer this question in section 3.6. However, the proof of theorem 3.31 required in an essential way the invariance of the attractors, at least for t ≥ 0. Since we only know that the exponential attractors are positively invariant for t ≥ 0, this procedure cannot be applied. On the other hand, the upper semicontinuity of the exponential attractors may follow from some recent results of Fabrie, Galinski, Miranville and Zelik, [FGMZ03], and of Gatti, Grasselli, Miranville and Pata, [GGMP03].

4.5 Proof of Theorem 4.4 In this section we prove theorem 4.4, following in large part the procedure presented in chapter 2 of Eden, Foias, Nicolaenko and Temam, [EFNT94].

4.5.1 Outline We refer to the second alternative characterization of the discrete squeezing property “(4.14) =⇒ (4.11)”. More precisely, we assume that the map S : X → X is Lipschitz continuous, that B is a nonempty, compact and positively invariant subset of X , that X admits an N-dimensional subspace XN , with corresponding orthogonal projection PN : X 7→ XN , and, finally, that there is some γ ∈ ]0, 12 [, with the property that, if u, v ∈ B are such that √ kS(u) − S(v)k ≥ 2kPN (S(u)) − PN (S(v))k , (4.52) then kS(u) − S(v)k ≤ γku − vk .

(4.53)

In the sequel, we omit the dependence on N for convenience, i.e. we write PN = P; also, we write Px instead of P(x), etc. We know from theorem 2.31 that A = ω(B) is the global attractor of S. We construct E as the union of A and a set generated by an iterative procedure, at each stage of which we add to A a set of points which are “bad”, in the sense that the orbits generating from these points do not converge exponentially to A. Since exponential convergence is a consequence of the strict contractivity of S, these points must lie in

4.5

153

Proof of Theorem 4.4

a set where (4.53) doesn’t hold. Then, the squeezing property asserts that (4.52) must hold for points in this set. Thus, at each stage, the fundamental step of each iteration consists in adding these “noncontractive” points to the set Sn (B), by covering it with a minimal number of balls of decreasing size. The number of these balls is seen to be not increasing with n, as a consequence of the squeezing property. This provides a control, at each step, of the size of the enlarged set obtained by the covering of Sn (B). The inductive limit of this process produces a set, which is shown to be the desired exponential attractor E. In this construction, we make essential use of the facts that the global attractor A is invariant, i.e. A = S(A), and the subset B is positively invariant, i.e. S(B) ⊆ B. We shall also see that the squeezing property implies the exponential convergence to E of the orbits which start at all other points of X . In addition, the rate of convergence of the orbits in this construction of the exponential attractor depends only on γ; more precisely, the constant k in (4.1) is only required to satisfy the bounds 0 < k < − ln(2γ) (recall that γ < 21 ). Finally, we remark that this construction of the exponential attractor is not the only possible one; for alternative constructions, see e.g. chapter 7 of [EFNT94].

4.5.2 The Cone Property Recalling definition (4.10) of the cone CP defined by the orthogonal projection P, we give DEFINITION 4.14 Let C ⊆ X . C satisfies the CONE PROPERTY if for all x, y ∈ C, √ kx − yk ≤ 2kPx − Pyk . (4.54) For instance, any singleton C = {x} trivially satisfies the cone property. The following is a more interesting example, which will be the basis of the construction of inertial manifolds in chapter 5. Example 4.15 The graph of a Lipschitz continuous function m : P(X ) → (I − P)(X ), with Lipschitz constant L ≤ 1, satisfies the cone property. That is, any set of the form C(m) := {x ∈ X : x = ξ + m(ξ ) , ξ ∈ P(X )} ,

(4.55)

with m satisfying the global Lipschitz condition km(ξ1 ) − m(ξ2 )k ≤ Lkξ1 − ξ2 k ,

0 < L ≤ 1,

satisfies the cone property. Indeed, if x = ξ + m(ξ ) and y = η + m(η) ∈ C(m), then, recalling that the projections are orthogonal, kx − yk2 = kξ − ηk2 + km(ξ ) − m(η)k2 ≤ (1 + L2 )kξ − ηk2

154

4

Exponential Attractors

≤ 2kξ − ηk2 = 2kPx − Pyk2 ; i.e., (4.54) holds. Note also that the sets C(m) defined in (4.55) are MAXIMAL with respect to the cone property (4.54), in the sense that if C 0 is another set satisfying the cone property, and C(m) ⊆ C 0 , then C(m) = C 0 . To show this, let x ∈ C 0 , and set y := Px + m(Px). Then y ∈ C(m) ⊆ C 0 , and Py = Px. Since C 0 satisfies the cone property, (4.54) implies √ kx − yk ≤ 2kPx − Pyk = 0 . (4.56) Hence, x = y ∈ C(m), so that C 0 ⊆ C(m). This motivates the following DEFINITION 4.16 Let C ⊆ Z ⊆ X . C is MAXIMAL with respect to the cone property (4.54) in Z if C satisfies the cone property and, whenever C ⊆ C 0 ⊆ Z and C 0 satisfies the cone property, C 0 = C. For example, given Z ⊆ X , each set C(m) ∩ Z, with C(m) defined in (4.55), is maximal with respect to the cone property in Z. We next show that we can always construct subsets of any given nonempty set, which are maximal with respect to the cone property in this set. PROPOSITION 4.17 Let Z ⊆ X , with Z 6= ∅. Choose z ∈ Z. Then the set C :=

[

{D ⊆ Z : z ∈ D, D has the cone property}

(4.57)

is maximal with respect to the cone property in Z. PROOF Obviously, the singleton {z} is in C ⊆ Z, since it trivially satisfies the cone property. Let C 0 ⊇ C be another subset of Z which satisfies the cone property. Then z ∈ C 0 and, therefore, C 0 is one of the sets in the union at the right side of (4.57). Hence, C 0 ⊆ C; therefore, C 0 = C. Thus, without loss of generality, we can consider sets which are maximal with respect to the cone property (4.54) in any given nonempty set Z. PROPOSITION 4.18 Let C ⊆ Z ⊆ X be maximal with respect to the cone property in Z, and let x ∈ X . Then x ∈ C if and only if there is y ∈ C such that y − x ∈ CP . PROOF Assume first that x ∈ C. Taking y = x, the cone property (4.54) is trivially satisfied. Conversely, assume there is y ∈ C such that x and y satisfy (4.54). Then x

4.5

Proof of Theorem 4.4

155

and y belong to a larger set C˜ ⊆ Z satisfying the cone property. Since C is maximal, C˜ = C; hence, x ∈ C. The following two results do not require C to be maximal, but only to satisfy the cone property. PROPOSITION 4.19 Let C ⊆ X satisfy the cone property. The orthogonal projection P is injective on C. PROOF This is an immediate consequence of (4.56). PROPOSITION 4.20 Let C ⊆ X satisfy the cone property. Then, C has no interior. PROOF Arguing by contradiction, assume the interior of C contains a point x. There is then r > 0 such that B(x, r) ⊂ C. Let y ∈ B(x, r) be such that y = 6 x and y 6= Py, and set r y − Py z := x + 2 ky − Pyk (if x 6= Px, we can take y = x). Then z ∈ B(x, r), since kz − xk = 2r . Thus, both x and z are in C, so by the cone property (4.54), kz − xk ≤

√

2kPz − Pxk =

√

√ r Py − P2 y r P(y − Py)

= 2

. − Px 2 Px +

2 ky − Pyk 2 ky − Pyk

Recalling that P2 = P, this implies a contradiction.

4.5.3 The Basic Covering Step In this section we describe the basic step in the construction of the covering of the global attractor A; this step, repeated inductively, will provide the basis for the construction of an exponential attractor E, containing A. We adopt the following notations: For a ∈ X and r > 0, B(a, r) is the closed ball with center a and radius r. As in (2.39) of section 2.7.1, diam(C) denotes the diameter of a subset C ⊂ X . In the remaining part of this section, we fix a compact set K ⊂ X and, given a ∈ K and r > 0 we define the set Z := S(K ∩ B(a, r)) . Since S is continuous and K is compact, Z is also compact.

(4.58)

156

4

Exponential Attractors

PROPOSITION 4.21 Let C ⊆ Z be a set which is maximal with respect to the cone property (4.54). Then C is compact in X . PROOF Since C is a subset of the compact set Z, it is sufficient to show that C is closed. Thus, let (xm )m∈N ⊂ C be such that xm → x. Since C ⊆ Z and Z is compact, x ∈ Z. By proposition 4.18, for each m ∈ N there is ym ∈ C such that xm − ym ∈ CP . Since (ym )m∈N ⊂ Z, there is a subsequence (ymk ) converging to a limit y ∈ Z as k → +∞. Since P is continuous, by (4.54) √ √ kx − yk = lim kxmk − ymk k ≤ 2 lim kPxmk − Pymk k = 2kPx − Pyk . k→+∞

k→+∞

This means that x − y ∈ CP . Thus, x and y belong to a larger set C˜ ⊆ Z satisfying the cone property. Since C is maximal with respect to the cone property, C˜ = C; hence, x ∈ C, and C is closed. Since C is compact and P is continuous, the set P(C) is also compact. Thus, given any ρ > 0, P(C) can be covered by a finite number K0 (ρ) of balls B(Py j , ρ), j = 1, . . . , K0 (ρ), y j ∈ C. Since Py j 6= Pyk (by proposition 4.19), we can obviously choose these balls so that kPy j − Pyk k > ρ

if j 6= k .

(4.59)

We now show that this covering of P(C) can be lifted to a covering of C, and to one of Z. We note explicitly that since C and Z are compact, they can obviously be covered by a finite number of balls; however, we want to cover C and Z with a specific covering, for which we have an explicit estimate on the radii of the covering balls (see proposition 4.24 below). PROPOSITION 4.22 Let C be as in proposition 4.21. Fix ρ > 0, and let y1 , . . . , yK0 (ρ) be as above. Then K0 (ρ)

C⊂

[

√ B(y j , 2ρ) .

j=1

PROOF Let y ∈ C. Since Py ∈ P(C), there is j ∈ {1, . . . , K0 (ρ)} such that Py ∈ B(Py j , ρ). Since both y and y j ∈ C, by (4.54) √ √ ky − y j k ≤ 2kPy − Py j k < 2ρ ; √ thus, y ∈ B(y j , 2ρ). Next, we see that this same covering of C generates a covering of Z.

4.5

157

Proof of Theorem 4.4

PROPOSITION 4.23 Given γ as in (4.53), r as in (4.58), and ρ > 0, set √ ρ1 := 2γr + 2ρ .

(4.60)

Then K0 (ρ)

Z⊂

[

B(y j , ρ1 ) .

j=1

PROOF Let z ∈ Z. If z ∈ C, proposition 4.22 applies, since choose y ∈ C. Then, by proposition 4.18, z − y ∈ / CP , i.e. √ kz − yk ≥ 2kPz − Pyk .

√

2ρ < ρ1 . If z ∈ / C, (4.61)

Since z, y ∈ Z = S(K ∩ B(a, r)), there are u and v ∈ B(a, r) such that z = S(u) and y = S(v). Then (4.61) implies (4.52) and, therefore, (4.53) holds. Then, kz − yk ≤ γku − vk ≤ γ diam(B(a, r)) ≤ 2γr .

(4.62) √ Since y ∈ C, by proposition 4.22 there is j ∈ {1, . . . , K0 (ρ)} such that y ∈ B(y j , 2ρ). Hence, (4.62) implies that √ kz − y j k ≤ kz − yk + ky − y j k ≤ 2γr + 2ρ = ρ1 , i.e. z ∈ B(y j , ρ1 ). We now proceed to give a first estimate on the number K0 (ρ) of the balls of radius ρ, which cover Z. PROPOSITION 4.24 Let γ, r and ρ be as in proposition 4.23, and N be the rank of P = PN . Then, N 2rL K0 (ρ) ≤ +1 , ρ where L denotes the Lipschitz constant of S. PROOF Note first that if we cover P(C) with other balls, whose centers are exactly at a distance ρ apart, and H is the number of these balls, then K0 (ρ) ≤ H. In fact, by (4.59) the distance between the centers of these other balls is less than that between the centers Py j of the “original” balls, so there must be more of the “new” balls. Consider then any such covering, i.e. H [ h=1

B(Pxh , ρ) ⊃ P(C) ,

158

4

Exponential Attractors

with centers Pxh ∈ P(C) such that kPxk − Px j k = ρ. Then for all j 6= k B(Px j , ρ2 ) ∩ B(Pxk , ρ2 ) = ∅ , for otherwise if there were a z in one such intersection, we would deduce the contradiction ρ = kPx j − Pxk k ≤ kPx j − zk + kz − Pxk k < ρ . This disjoint family is also contained in the 12 ρ-neighborhood of P(C), i.e. the set V :=

H [

B(Pxh , ρ2 ) ⊆ {y ∈ P(X ) : d(y, P(C)) < ρ2 } =: U .

(4.63)

h=1

In fact, if x ∈ V, there is j ∈ {1, . . . , H} such that x ∈ B(Px j , ρ2 ); hence, d(x, P(C)) ≤ kx − Px j k <

ρ 2

.

Recalling that C ⊂ Z, from (4.63) we have the estimate diam(U) ≤ diam(P(C)) + ρ ≤ diam(P(Z)) + ρ ≤ diam(Z) + ρ

(4.64)

(the last step because P, being orthogonal, is contractive). To estimate diam(Z), let z1 and z2 ∈ Z. Since Z = S(K ∩ B(a, r)), there are b1 and b2 ∈ K ∩ B(a, r) such that zi = S(bi ), i = 1, 2. Recalling that S is Lipschitz continuous with Lipschitz constant L, we have kz1 − z2 k ≤ Lkb1 − b2 k ≤ 2rL . Consequently, we deduce that diam(Z) ≤ L diam(B(a, r)) = 2rL . Inserting this into (4.64) we deduce that diam(U) ≤ 2rL + ρ, which means that U is contained in a ball of radius rL + ρ2 . Since P(X ) is N-dimensional, we conclude that vol(U) ≤ ωN rL + ρ2

N

,

(4.65)

where ωN denotes the volume of the unit ball of RN . Now, the family V is composed N of H balls, each of which has volume ωN ρ2 . Hence, (4.63) and (4.65) imply that vol(V) = HωN

ρ N 2

≤ vol(U) ≤ ωN rL + ρ2

N

.

From this we conclude that H≤

2rL ρ

N +1 .

Since K0 (ρ) ≤ H, proposition 4.24 follows.

(4.66)

4.5

Proof of Theorem 4.4

159

We now show that we can choose a particular value of ρ, which allows us to estimate the number of the balls of the corresponding covering of C and Z independently of r. Indeed, given θ ∈ ]2γ, 1[, recalling definition (4.60) we can choose ρ such that ρ1 = θ r. More precisely, we choose ρ=

(θ − 2γ)r √ . 2

(4.67)

With this value of ρ, estimate (4.66) implies that K0 (ρ) ≤

!N √ 2 2L +1 . θ − 2γ

In conclusion, we can summarize the results of this section in the following THEOREM 4.25 Let K ⊂ X be a compact set. Given any a ∈ K and r > 0, set Z := S(B(a, r) ∩ K). There exists a compact subset C ⊆ Z ⊆ S(K) such that C is maximal with respect to the cone property (4.54). Moreover, for any θ ∈ ]2γ, 1[ there is a covering of both C and Z, consisting of balls of radius θ r and center in C. This covering has the form C⊆Z ⊂

M [

yj ∈ C ,

(4.68)

!N √ 2 2L +1 . θ − 2γ

(4.69)

B(y j , θ r) ,

j=1

and consists of M balls, where M ≤ H1 (θ ) :=

PROOF The existence and compactness of the set C follow from propositions 4.17 and 4.59. Given then θ ∈ ]2γ, 1[, choosing ρ as in (4.67) we have from (4.60) that ρ1 = θ r. With this value of ρ, estimate (4.66) implies (4.69). Noting that the right side of (4.69) is independent of both r and ρ, we conclude that there exists a covering as claimed in (4.68).

4.5.4 The First and Second Iterates Let B ⊂ X be a compact, positively invariant set for the semiflow S, containing the global attractor A. By repeatedly iterating the construction described in the previous section, we define successive coverings of the sets Sk (B), for integer k ≥ 0, in a way that allows us to keep a suitable control on the radii of the balls covering Sk (B).

160

4

Exponential Attractors

More precisely, we want the radius of each ball of the covering of Sk (B) to be equal to θ k R, where R > 0 and θ ∈ ]2γ, 1[ are fixed. (A specific choice of the radius R will be determined later on, when we prove the estimate on the fractal dimension of E.) The starting point is the covering of B itself. Since B is compact, we can cover it by a finite number M0 of balls of radius 1; that is, there are points a1 , . . . , aM0 in B such that B⊆

M0 [

B(a j , 1) .

j=1

Obviously, then, for any R ≥ 1, M0 [

B⊆

(B(a j , R) ∩ B) .

(4.70)

j=1

This is the covering step corresponding to k = 0; to emphasize this, we rewrite (4.70) as B⊆

M0 [

(B(a j0 , θ 0 R) ∩ S0 (B)) .

(4.71)

j0 =1

We remark that, since B is compact, we could cover it by a single ball of a sufficiently large radius R. We use the covering (4.70) instead, because in the estimate of the fractal dimension of E we will need to be able to take R as close to 1 as we want (see step (8) of the proof of theorem 4.26 below). For the next step, i.e. the covering of S(B), we first deduce from (4.71) that S(B) ⊆

M0 [

S(B(a j0 , θ 0 R) ∩ S0 (B)) .

(4.72)

j0 =1

Then, we use theorem 4.25, with r = R and K = S0 (B) = B (which is compact), to cover each set at the right side of (4.72). More precisely, setting (1)

B j0 := S(B(a j0 , θ 0 R) ∩ S0 (B)) , (1)

we can cover each set B j0 , 1 ≤ j0 ≤ M0 , with M1, j0 balls, M1, j0 ≤ H1 (θ ) (defined in (1)

(1)

(4.69)), having radii θ R and centers in a subset E j0 of B j0 , maximal with respect to (1)

(1)

the cone property (4.54) (that is, B j0 and E j0 are, respectively, the sets Z and C of (1)

theorem 4.25). Explicitly, this means that there are points a j0 j1 ∈ E j0 , 1 ≤ j1 ≤ M1, j0 , such that (1) B j0

M1, j0

⊂

[ j1 =1

B(a j0 j1 , θ R) .

4.5

161

Proof of Theorem 4.4

From this and (4.71), it follows that 1

S (B) ⊂

M

M0 [

(1) B j0

⊂

j0 =1

M0 1, j0 [ [

B(a j0 j1 , θ R)

j0 =1 j1 =1 M

⊂

M0 1, j0 [ [

(B(a j0 j1 , θ R) ∩ S1 (B)) .

(4.73)

j0 =1 j1 =1

This describes the second step of our construction, corresponding to k = 1. In the (1) terminology of [EFNT94], the sets E j0 , 1 ≤ j0 ≤ M0 , make up the “first generation” of the points to be added to the attractor A.

4.5.5 The General Iterate We now proceed inductively. In analogy to the previous construction, we set (0)

E j0 , j−1 := B and, for k ≥ 0, define the compact set (k+1)

B j0 ... jk := S(B(a j0 ... jk , θ k R) ∩ Sk (B)) ,

(4.74)

where the balls B(a j0 ... jk , θ k R) have centers (k)

a j0 ... jk ∈ E j0 ... jk−1 ,

(4.75)

(k+1)

and E j0 ... jk is the compact set C, maximal with respect to the cone property (4.54) (k+1)

in the set Z = B j0 ... jk , corresponding, in theorem 4.25, to the choices r = θ k R and K = Sk (B) (which, being the continuous image of a compact set, is compact). Thus, for all k ≥ 0, (k+1)

(k+1)

E j0 ... jk ⊆ B j0 ... jk ,

(4.76)

and both these sets are compact. To proceed with the iteration, we apply again the(k+1) orem 4.25 to cover the set B j0 ... jk defined in (4.74) by Mk+1, j0 ... jk balls with centers (k+1)

a j0 ... jk+1 ∈ E j0 ... jk and radii θ k+1 R, with Mk+1, j0 ... jk ≤ H1 (θ )

(4.77)

(defined in (4.69)); that is, (k+1) B j0 ... jk

Mk+1, j0 ... j

⊂

[ jk+1 =1

k

B(a j0 ... jk+1 , θ k+1 R) .

(4.78)

162

4

Exponential Attractors

We now show that this covering satisfies the inclusions E (k+1) :=

Mk, j0 ... j

M0 [

[ k−1 (k+1)

E j0 ... jk ⊂ Sk+1 (B) ,

···

j0 =1

S

k+1

(B) ⊂

Mk, j0 ... j

M0 [

[ k−1 (k+1)

···

B j0 ... jk

j0 =1

⊂

(4.79)

jk =1

jk =1 Mk, j0 ... j

M0 [

[ k−1

···

Mk+1, j0 ... j

jk =1

j0 =1

[

k

B(a j0 ... jk+1 , θ k+1 R) ∩ Sk+1 (B) .

(4.80)

jk+1 =1

To show (4.79), let x ∈ E (k+1) . There are then indices j0 ∈ {1, . . . , M0 }, . . ., jk ∈ (k+1) {1, . . . , Mk, j0 ... jk−1 } such that x ∈ E j0 ... jk . Thus, by (4.76) and (4.74), x ∈ S(B(a j0 ... jk , θ k R) ∩ Sk (B)). Consequently, there is y ∈ B(a j0 ... jk , θ k R) ∩ Sk (B) such that x = S(y); since y ∈ Sk (B), it follows that x ∈ Sk+1 (B). To show (4.80), we proceed by induction on k. The step k = 0 is (4.73). Then, if (4.80) holds with k replaced by k − 1, k ≥ 1, by (4.74) we have   M S

k+1

k

(B) = S(S (B)) ⊆ S 

M0 [

k, j0 ... jk−1

···

j0 =1

⊂

M0 [

Mk, j0 ··· j

···

j0 =1

=

M0 [ j0 =1

[

(B(a j0 ... jk , θ k R) ∩ Sk (B))

jk =1

[ k−1

S B(a j0 ··· jk , θ k R) ∩ Sk (B)

jk =1 Mk, j0 ··· j

···

[ k−1 (k+1)

B j0 ··· jk .

jk =1

Thus, (4.80) follows by (4.78). Finally, note that each set E (k+1) , being a finite union of compact sets, is itself compact.

4.5.6 Conclusion We are almost at the end of the construction of the exponential attractor in B. We recall that we do have a global attractor A in B, namely the set A = ω(B). The sets E (k) constructed in (4.79) are defined in terms of sets C where the cone property holds: this means that the semiflow S needs not be contracting on these sets. Hence, we add all these points to the attractor A. More precisely, setting E (∞) : =

∞ [

k=1

E (k)

(4.81)

4.5

Proof of Theorem 4.4

163

(closure in X ), and E1 : = A ∪ E (∞) , we would like to recognize E1 as a “good candidate” for the exponential attractor. Indeed, note that E1 is closed, because so are A and E (∞) , and that E1 ⊆ B, because A ⊆ B and also E (k) ⊆ B for all k ∈ N. The latter claim is a consequence of (4.79) and of the positive invariance of B, which implies that E (k) ⊂ Sk (B) ⊂ B .

(4.82)

Since B is compact, it follows that E (∞) and E1 are also compact. However, E1 needs not be positively invariant. To overcome this last problem, we enlarge E1 by taking all its images under the iterates of S. More precisely, setting G (∞) : =

∞ [

S j (E (∞) ) , E := A ∪ G (∞) ,

(4.83)

j=0

we finally claim THEOREM 4.26 The set E defined in (4.83) is an exponential attractor for the semiflow S = (Sn )n∈N , relative to B. PROOF Recalling definition 4.1, we must show that E ⊆ B; that E is positively invariant and compact; that E attracts bounded subsets of B at an exponential rate; and that E has finite fractal dimension. Note that, by construction, A ⊆ E. 1. We show that E ⊆ B. Indeed, A ⊆ B and, since E (∞) ⊆ B (which follows from (4.82)), for all j ∈ N S j (E (∞) ) ⊆ S j (B) ⊆ B . Therefore, G (∞) ⊆ B as well. 2. In preparation for the next steps, for j ≥ 0 we set L( j) : =

∞ [

S j (E (k) ) ,

k=1

and define L(∞) : =

∞ [

j=0

L( j) , E2 := A ∪ L(∞) .

(4.84)

164

4

Exponential Attractors

We prove then that E = E2 , and proceed to show that E2 satisfies the requirements of the exponential attractor. 3. We show that E = E2 . 3a. To show that E ⊆ E2 , let x ∈ E. If x ∈ A, there is nothing more to prove. If x ∈ G (∞) , there is j ∈ N such that x ∈ S j (E (∞) ); that is, x = S j (y) for some y ∈ E (∞) . Let (ym )m∈N ⊂

∞ [

E (k)

k=1

be a sequence converging to y. Then for all m there is km such that ym ∈ E (km ) . By (4.79), ym ∈ Skm (B), so ym = Skm (bm ) for some bm ∈ B. If the sequence (km )m∈N is unbounded, recalling the characterization of ω-limit sets given in proposition 2.15, we conclude that y = lim ym ∈ ω(B) = A . Consequently, since A is invariant, x = S j (y) ∈ S j (A) = A ⊆ E2 . If instead (km )m∈N is definitely constant, i.e. if there are m0 and k∗ ∈ N such that km ≡ k∗ for all m ≥ m0 , then the sequence (ym )m≥m is all contained in the compact 0

set E (k∗ ) . Thus, there is a subsequence (ymr )r∈N , with mr ≥ m0 , converging to some y∗ ∈ E (k∗ ) . This implies that y = y∗ ∈ E (k∗ ) , and, therefore, x = S j (y) ∈ S j (E (k∗ ) ) ⊆ L( j) ⊆ L∞ ⊆ E2 . It follows that E ⊆ E2 . 3b. To show that E2 ⊆ E, let x ∈ E2 . If x ∈ A, there is nothing more to prove. If x ∈ L(∞) , there are integers j and k such that x ∈ S j (E (k) ), so x = S j (y) for some y ∈ E (k) . But E (k) ⊆ E (∞) , so x ∈ S j (E (∞) ) ⊆ G (∞) ⊆ E . 4. We show that E2 is positively invariant. Let x ∈ E2 . If x ∈ A, then S(x) ∈ S(A) = A ⊆ E2 . If x ∈ L(∞) , as before there are integers j and k such that x ∈ S j (E (k) ). Then S(x) ∈ S j+1 (E (k) ) ⊆ L( j+1) ⊆ L(∞) ⊆ E2 , and, therefore, S(E2 ) ⊆ E2 . Note that it is precisely to ensure the positive invariance of E that we had to enlarge E1 into E2 . 5. We show that E2 is closed. Let (am )m∈N ⊂ E2 be a sequence such that am → a, a ∈ X . If there is a proper subsequence (amk )k∈N ⊂ A, then a = lim amk ∈ A ⊂ E2 . k→+∞

4.5

Proof of Theorem 4.4

165

Otherwise, at most a finite number of elements of (am )m∈N are in A. We reorder the other elements into a sequence (ym )m∈N ⊂ L(∞) , with ym → a. Since each ym ∈ L(∞) , for all m there exist nm and km such that ym = Snm (xm ) , xm ∈ E (km ) ⊆ B .

(4.85)

Suppose first that the sequence (nm )m∈N is unbounded. Then by the already recalled characterization of A = ω(B) given in proposition 2.15, (4.85) implies that a = lim ym ∈ ω(B) = A ⊂ E2 . m→+∞

If instead the sequence (nm )m∈N is definitely constant, i.e. if there are m0 and n∗ ∈ N such that for all m ≥ m0 , nm = n∗ , then for m ≥ m0 ym = Sn∗ (xm ) . Since (xm )m∈N ⊂ B and B is compact, there is at least a subsequence (xm j ) j∈N converging to some x ∈ B. Since Sn∗ is continuous, we have that for all j such that m j ≥ m0 ym j = Sn∗ (xm j ) → Sn∗ (x) , and this implies that a = Sn∗ (x). Consider now the subsequence (km j ) j∈N of the sequence (km )m∈N appearing in (4.85), and suppose first that this sequence is definitely constant, i.e., that there are j0 and k∗ ∈ N such that for all j ≥ j0 , km j = k∗ . Then for all j ≥ j0 , (k ) xm j ∈ E m j = E (k∗ ) , and since this set is compact, x = lim xm j ∈ E (k∗ ) . j→+∞

It follows that a = Sn∗ (x) ∈ Sn∗ (E (k∗ ) ) ⊆ L(n∗ ) ⊆ L(∞) ⊆ E2 . The last possibility is that km j → +∞ as j → +∞. Then, as before, xm j ∈ E

(km j )

⊆S

km j

(B) ⊆ B ,

and therefore, by proposition 2.15 again, x ∈ ω(B) = A, and a = Sn∗ (x) ∈ Sn∗ (A) = A ⊂ E2 . This concludes the proof that E2 is closed. 6. We show that E2 is compact. Since B is compact and E2 is closed, it is sufficient to show that E2 ⊆ B. But this follows from steps 1 and 3, since we know that E ⊆ B and E2 = E. 7. We show the exponential convergence of bounded subsets of B to E2 . More precisely, setting κ := − ln θ ∈ ]0, − ln(2γ)[

166

4

Exponential Attractors

(recall that 2γ < 1), we show that for any bounded subset G ⊆ B, the estimate ∂ (Sn G, E2 ) ≤ Re−κn

(4.86)

holds for all n ∈ N (compare to (4.1)). To this end, we recall (4.80), which implies that for all x ∈ G and n ∈ N, n

n

n

S (x) ∈ S G ⊆ S (B) ⊂

M0 [

Mn, j1 ··· jn−1

···

j0 =1

[

B(a j0 ··· jn , θ n R ∩ Sn (B)).

jn =1

Thus, there are indices j0 , . . . , jn such that Sn (x) ∈ B(a j0 ··· jn , θ n R) . Now, each center a j0 ··· jn is in E2 , because by (4.75) and (4.79), (n)

a j0 ··· jn ∈ E j0 ··· jn−1 ⊆ E (n) = S0 (E (n) ) ⊆ L(0) ⊆ L(∞) ⊆ E2 . Consequently, d(Sn (x), E2 ) ≤ kSn (x) − a j0 ··· jn k ≤ θ n R , and (4.86) follows. This concludes the proof of the exponential rate of convergence to E of all orbits starting in G. 8. We show that E has finite fractal dimension, satisfying estimate (4.15). To this end, we use the first of part (3) of proposition 2.61, applied to the compact sets A and G ∞ (the compactness of G ∞ follows from the fact that both sets A and A ∪ G ∞ = E = E2 are compact). 8a. We first estimate the fractal dimension of the global attractor A. We claim that dimF (A) ≤

ln H1 (θ ) , − ln θ

(4.87)

where θ ∈ ]2γ, 1[, and H1 (θ ) is as in (4.69). The proof of (4.87) is based on the observation that the same construction of the coverings of the sets Sk (B) described in sections 4.5.4 and 4.5.5 can be carried out, in the same way, when the set B is replaced by the attractor A itself, since A is also compact and positively invariant. But since A is in fact invariant, (4.80) shows that for each positive integer k, the set A = Sk (A) can be covered by M˜ k balls of radius θ k R, where M˜ k := M0 M1, j0 · · · Mk, j1 ,..., jk−1 .

(4.88)

Since Nδ (A) is the minimum number of balls of diameter at most equal to δ that can cover A, recalling (4.77) it follows from (4.88) that N2θ k R (A) ≤ M˜ k ≤ M0 (H1 (θ ))k .

(4.89)

Given then δ > 0, let k ∈ N be such that 2θ k+1 R < δ ≤ 2θ k R .

(4.90)

4.5

167

Proof of Theorem 4.4

Then, since the function δ 7→ Nδ (A) is decreasing, we obtain from (4.89) and (4.90) that ln Nδ (A) ln N2θ k+1 R (A) (k + 1) ln H1 (θ ) + ln M0 ≤ ≤ . − ln δ − ln(2R) − k ln θ − ln(2θ k R) Hence, recalling (4.90), dimF (A) = lim sup δ →0

ln Nδ (A) (k + 1) ln H1 (θ ) + ln M0 ≤ lim , k→+∞ − ln δ − ln(2R) − k ln θ

from which (4.87) follows. 8b. As an intermediate step, we estimate the fractal dimension of E (∞) . We claim that n ln H (θ ) o 1 =: η0 , (4.91) dimF (E (∞) ) ≤ max N, − ln θ where N is the rank of the projection P appearing in definition 4.2, θ ∈ ]2γ, 1[ and H1 (θ ) is as in (4.69). To prove (4.91), we consider the coverings of the iterates Sk (B) constructed in sections 4.5.5, with R specified as follows. Given δ ∈ ]0, 1[, we first choose n ∈ N such that 2θ n+1 < δ ≤ 2θ n ,

(4.92)

2θ n+1 R = δ .

(4.93)

and then R > 0 such that

Note that both n and R depend on δ , but (4.92) implies that, for all δ ∈ ]0, 1[, 1
1 . θ

(4.94)

We proceed then to construct the sets E (k) as in section 4.5.5, with R as in (4.93). Recalling the definition of E (∞) in (4.81), we split ! ! n [

E (∞) =

E (k) ∪

k=1

|

{z } | =: E3

∞ [

E (k)

(4.95)

k=n+1

{z =: E4

}

(the first union E3 is automatically closed, being a finite union of compact sets); therefore, Nδ (E (∞) ) ≤ Nδ (E3 ) + Nδ (E4 ) . Recalling (4.79), the positive invariance of B implies that, if k ≥ n + 1, E (k) ⊆ Sk (B) ⊆ Sn+1 (B) .

(4.96)

168

4

Exponential Attractors

Since Sn+1 (B) is compact, it follows that E4 ⊆ Sn+1 (B), so that, by part 2 of proposition 2.61, Nδ (E4 ) ≤ Nδ (Sn+1 (B)).

(4.97)

To estimate Nδ (Sn+1 (B)), it is sufficient to recall (4.80), which implies that Sn+1 (B) can be covered by κ1 balls of radii θ n+1 R, with κ1 := M0 M1, j0 · · · Mn+1, j0 ··· jn . Thus, as in (4.89), and recalling (4.93), Nδ (Sn+1 (B)) = N2θ n+1 R (Sn+1 (B)) ≤ M0 (H1 (θ ))n+1 .

(4.98)

We now turn to the estimate of Nδ (E3 ). Recalling (4.79) and (4.80), we have E (k+1) ⊂

M0 [

Mk+1, j0 ··· j

k

[

···

j0 =1

B(a j0 ··· jk+1 , θ k+1 R) ∩ Sk+1 (B) ,

(4.99)

jk+1 =1

(k+1)

with a j0 ··· jk+1 ∈ E j0 ··· jk as per (4.75). We now project this covering of E (k+1) into the N-dimensional space P(X ). Denoting by BN (a, r) the closed balls of P(X ) with center a and radius r, we immediately verify that for all a ∈ X , r > 0 and m ∈ N, P(B(a, r) ∩ Sm (B)) ⊆ BN (Pa, r) ∩ P(Sm (B)) ;

(4.100)

hence, we obtain from (4.99) that P(E

(k+1)

)⊂

Mk+1, j0 ··· j

M0 [

[

···

j0 =1

k

BN (Pa j0 ··· jk+1 , θ k+1 R) ∩ P(Sk+1 (B)) .

(4.101)

jk+1 =1

We further cover each intersection in the sets at the right side of (4.101) by ν j0 ··· jk+1 δ smaller balls of radius 2√ (these balls are truly smaller, because of (4.93), and k ≤ n). 2 More precisely, there are points ξ` j

0 ··· jk+1

∈ BN (Pa j0 ··· jk+1 , θ k+1 R) ∩ P(Sk+1 (B)) ,

with ξ` j

0 ··· jk+1

= Py` j

0 ··· jk+1

(k+1)

, y` j

0 ··· jk+1

∈ E j0 ··· jk ,

(4.102)

such that BN (Pa j0 ··· jk+1 , θ k+1 R) ∩ P(Sk+1 (B)) ν j0 ··· jk+1

⊆

k+1 1 BN ξ` j ··· j , 2√ δ ∩ P(S (B)) . 2 0 k+1 {z } ` j0 ··· jk+1 =1 | =: D`∗j ··· j [

0

k+1

(4.103)

4.5

169

Proof of Theorem 4.4

The second of (4.102) holds, because the balls B(a j0 ··· jk+1 , θ k+1 R), which cover (k+1)

(k+1)

B j0 ··· jk (by (4.78)), have centers in E j0 ··· jk (as per (4.75)). In fact, these balls are (k+1)

a lifting of a covering of E j0 ··· jk , as seen in proposition 4.23 (recall that, at each stage (k+1)

(k+1)

k, we identify C = E j0 ··· jk and Z = B j0 ··· jk ). Therefore, from (4.101), P(E

(k+1)

)⊂

M0 [

···

Mk+1, j0 ··· j ν j0 ··· jk+1 [ k [

j0 =1

jk+1 =1

D`∗j

` j0 ··· jk =1

0 ··· jk+1

.

(4.104)

Since dim(P(X )) = N, we can estimate the number ν j0 ··· jk+1 in terms of N and the δ , θ n+1 R, as follows. Recall that, in general, if a ball BN (a, ρ) is covered by radii 2√ 2 ν balls with centers b1 , . . . , bν ∈ BN (a, ρ) and smaller radius η, from the inclusions BN (a, ρ) ⊆

ν [

BN (b j , η) =: U ⊆ BN (a, ρ + η)

j=1

we deduce that, in particular, vol(BN (a, ρ)) = ωN ρ N ≤ vol(U) = νωN η N ≤ vol(BN (a, ρ + η)) = ωN (ρ + η)N . Thus, we have the estimate

ν ≤ 1+

ρ η

N

N

N j N ρ =∑ ≤ ηρ η j j=0

N

N N ∑ j = 2N ηρ . j=0

(4.105)

When applied to η=

1 √ δ 2 2

< ρ = θ k+1 R ,

(4.105) yields the estimate ν j0 ··· jk+1 ≤ CN θ (k−n)N , with CN := 2N/2 . Hence, recalling (4.104), P(E (k+1) ) can be covered by κ2,k balls of δ radius 2√ , with, as in (4.89), 2 κ2,k : = M0 M1, j0 · · · Mk+1, j0 ··· jk ν j0 ··· jk+1 ≤ CN M0 (H1 (θ ))k+1 θ (k−n)N .

(4.106)

We now show that the covering (4.104) of P(E (k+1) ) can be lifted to a covering of itself, by means of the cone property (4.54). More precisely, we show that, if the points y` j ··· j are defined as in (4.102), then E (k+1)

0

E

(k+1)

k+1

⊂

M0 [ j0 =1

Mk+1, j0 ··· j

···

[ jk+1 =1

k

ν j0 ··· jk+1

[ ` j0 ··· jk+1 =1

B(y` j

0 ··· jk+1

, θ n+1 R) .

(4.107)

170

4

Exponential Attractors (k+1)

In fact, let y ∈ E (k+1) . By (4.79), there are indices j0 · · · , jk such that y ∈ E j0 ,..., jk . By (4.76) and (4.78), there is another index jk+1 such that y ∈ B(a j0 ,..., jk+1 , θ k+1 R). Indeed, by (4.99), y ∈ B(a j0 ··· jk+1 , θ k+1 R) ∩ Sk+1 (B) . By (4.100) and (4.103), there is one index ` j0 ··· jk+1 ∈ {1, . . . , ν j0 ··· jk+1 } such that 1 δ ∩ P(Sk+1 (B)) . Py ∈ BN ξ` j ··· j , 2√ 2 0

k+1

(k+1)

Recalling (4.102), and that the set E j0 ··· jk is maximal with respect to the cone property (4.54), we have √ √ ky − y` j ··· j k ≤ 2kPy − Py` j ··· j k = 2kPy − ξ` j ··· j k 0 0 0 k+1 k+1 k+1 √ 1 n+1 δ = θ R . ≤ 2 2√ 2 Thus, (4.107) follows. Recalling (4.106), we conclude that E (k+1) can be covered by κ2,k balls of diameter 2θ n+1 R = δ . Consequently, (4.106) implies that Nδ (E (k+1) ) ≤ CN θ (k−n)N (H1 (θ ))k+1 . This estimate holds for 0 ≤ k ≤ n − 1; hence, recalling (4.95), n

Nδ (E3 ) ≤

n

∑ Nδ (E (k) ) ≤ CN

∑ θ (k−1−n)N (H1 (θ ))k

k=1

k=1 n

k ≤ CN θ −N(1+n) ∑ θ N H1 (θ ) .

(4.108)

k=1

Putting (4.96) and (4.108) together with (4.97) and (4.97) and (4.98), we finally obtain that n k Nδ (E (∞) ) ≤ CN θ −N(1+n) ∑ θ N H1 (θ ) + M0 (H1 (θ ))n+1 .

(4.109)

k=1

To estimate the right side of (4.109), suppose first that θ N H1 (θ ) ≤ 1. In this case, we can proceed from (4.109) with Nδ (E (∞) ) ≤ CN θ −N(1+n) n + M0 (H1 (θ ))n+1 ≤ θ −N(1+n) (CN (n + 1) + M0 ) . By (4.93), θ n+1 =

δ ; 2R

thus, n ≤ n+1 =

ln(δ /2R) , ln θ

(4.110)

4.5 and Nδ (E (∞) ) ≤

171

Proof of Theorem 4.4

2R δ

N ln(δ /2R) CN + M0 . ln θ

Therefore, since δ < 1, 1 ln(δ /2R) ln(Nδ (E (∞) )) N(ln(2R) − ln δ ) ≤ + ln CN + M0 − ln δ − ln δ − ln δ ln θ =: h1 (δ ) .

(4.111)

Recalling that, by (4.94), R is a bounded function of δ , we conclude from (4.111) that, if θ N H1 (θ ) ≤ 1, dimF (E (∞) ) = lim sup δ →0

ln(Nδ (E (∞) )) ≤ lim h1 (δ ) = N . − ln δ δ →0

(4.112)

If instead θ N H1 (θ ) > 1, we proceed from (4.109) with Nδ (E (∞) ) ≤ CN θ −N(1+n) n(θ N H1 (θ ))n + M0 (H1 (θ ))n+1 ≤ CN θ −N n(H1 (θ ))n + M0 (H1 (θ ))n+1 ≤ CN n(H1 (θ ))1+n + M0 (H1 (θ ))n+1 = (H1 (θ ))n+1 (CN n + M0 ). From this and (4.110) we obtain that ln(Nδ (E (∞) )) (ln δ − ln(2R)) ln(H1 (θ )) 1 ln(δ /2R) ≤ + ln CN ln θ + M0 − ln δ (− ln δ )(ln θ ) − ln δ =: h2 (δ ) . We can then conclude, as in (4.112), that dimF (E (∞) ) ≤ lim h2 (δ ) = δ →0

ln(H1 (θ )) . − ln θ

(4.113)

Thus, (4.91) follows from (4.112) and (4.113). 8c. Our last step is the estimate of the fractal dimension of G (∞) . We claim that also dimF (G (∞) ) ≤ η0 ,

(4.114)

where η0 is as in (4.91). We prove (4.114) by showing that for all η > η0 , dimF (G (∞) ) ≤ η .

(4.115)

As in part (8b), given δ ∈ ]0, 1[ and θ ∈ ]2γ, 1[, we determine n ∈ N and R ∈ ]1, θ1 [ such that (4.92) and (4.93) hold. Then, as in (4.95), recalling (4.83) we split ! ! n [

G (∞) =

S j (E (∞) ) ∪

j=0

|

∞ [

S j (E (∞) )

,

j=n+1

{z =: E5

} |

{z =: E6

}

172

4

Exponential Attractors

so that Nδ (G (∞) ) ≤ Nδ (E5 ) + Nδ (E6 ) .

(4.116)

From (4.84) and the positive invariance of B we have that, if j ≥ n + 1, S j (E (∞) ) ⊆ S j (B) ⊆ Sn+1 (B) ; hence, E6 ⊆ Sn+1 (B) and, therefore, Nδ (E6 ) ≤ Nδ (Sn+1 (B)) .

(4.117)

We can estimate Nδ (Sn+1 (B)) exactly as in (4.98), with the number M0 replaced by M˜ 0 := N2 (E (∞) ) , i.e. by the minimum number of balls of radius 1 that are needed to cover the compact set E (∞) . Thus, from (4.117), Nδ (E6 ) ≤ Nδ (Sn+1 (B)) ≤ M˜ 0 (H1 (θ ))n+1 .

(4.118)

We now turn to the estimate of Nδ (E5 ). To this end, given an index j ∈ {0, . . . , n}, we first cover E (∞) by exactly Nδ θ − j (E (∞) ) balls of radius −j 1 2δθ

= θ n+1− j R .

Then, by repeated applications of theorem 4.25, with the choice, at each stage k, 0 ≤ k ≤ j, of r = θ n+1− j+k R and K = Sk (E (∞) ), we obtain that the set S j (E (∞) ) can be covered by β j balls of radius θ n+1 R = 12 δ . For example, the step k = 0 is as follows. Letting r0 := θ n+1− j R and ν0 := Nδ θ − j (E (∞) ), the chosen covering of E (∞) yields the inclusion β0 [

E (∞) ⊆

(E (∞) ∩ B(ai0 , r0 )) ,

i0 =1

with suitable centers a1 , . . . , aβ0 ∈ E (∞) . Thus, S(E (∞) ) ⊆

β0 [

S(E (∞) ∩ B(ai0 , r0 )) ,

(4.119)

i0 =1

and theorem 4.25 implies that each set at the right side of (4.119) can in turn be covered by ν1 balls of radius θ r0 = θ n+1− j+1 R. This yields the step for k = 1. Proceeding in this way, we finally cover S j (E (∞) ) by ν0 ν1 · · · ν j =: β j balls of radius θ j r0 = θ n+1 R. As in (4.89), β j can be estimated by β j ≤ Nδ θ − j (E (∞) )(H1 (θ )) j ;

4.5

173

Proof of Theorem 4.4

since Nδ (S j (E (∞) )) ≤ β j , we eventually obtain that Nδ (E5 ) ≤

n

n

j=0

j=0

∑ Nδ (S j (E (∞) )) ≤ ∑ Nδ θ − j (E (∞) )(H1 (θ )) j .

(4.120)

Let now η0 be as in (4.91). We claim that for all η > η0 , there exists cη ≥ 1 such that for all ε ≤ 1, η Nε (E (∞) ) ≤ cη ε1 . (4.121) Indeed, by the very definition of the limes superior, i.e., the lowest upper bound, (4.91) implies that for all η > η0 there is an ε0 > 0 such that for all ε ≤ ε0 , ln(Nε (E (∞) )) ≤η; − ln ε

(4.122)

we can clearly restrict ε0 to be in ]0, 1[. Then: if ε ≤ ε0 , (4.122) implies that η Nε (E (∞) ) ≤ ε1 ; if instead ε0 < ε ≤ 1, Nε (E (∞) ) ≤ Nε0 (E (∞) ) ≤

η 1 ε0

≤

η 1 ε0

1 η ε

=: cη

1 η ε

,

(4.123)

with cη := ε0−η depending (only) on η via ε0 , as per (4.122). Since cη ≥ 1, (4.122) and (4.123) show that (4.121) holds. Given then δ and j as above, we specify ε as ε = 12 θ − j δ =: ε j , and note that ε j ≤ 1 for all j ∈ {0, . . . , n}, because of (4.93) and (4.94). Hence, (4.121) implies that, for 0 ≤ j ≤ n, j η . Nδ θ − j (E (∞) ) ≤ Nε j (E (∞) ) ≤ cη 2θδ Thus, we obtain from (4.120) that n

Nδ (E5 ) ≤ cη

∑

2θ j δ

η

(H1 (θ )) j = cη

j=0

2 η δ

n

∑ (θ η H1 (θ )) j .

(4.124)

j=0

As in part (8b), we now distinguish two cases. Assume at first that θ η H1 (θ ) ≤ 1. Then from (4.116), (4.118) and (4.124) we have η (4.125) Nδ (G (∞) ) ≤ cη δ2 (n + 1) + M˜ 0 (H1 (θ ))n+1 . We now recall that, from (4.93), n+1 =

ln(δ /2R) ln(2R/δ ) = ; ln θ − ln θ

174

4

Exponential Attractors

thus, using the identity xln y = yln x (x, y > 0), we obtain from (4.125) that η 1/(− ln θ ) 2 ln(2R/δ ) (∞) Nδ (G ) ≤ cη + M˜ 0 (H1 (θ ))ln(2R/δ ) δ − ln θ η Θ cη 2R 2R ln(2R/δ ) = η + M˜ 0 , (4.126) R δ − ln θ δ where Θ :=

ln(H1 (θ )) . − ln θ

Since Θ ≤ η0 ≤ η, (4.126) implies that η cη ln(2R) − ln δ 2R ˜0 . Nδ (G (∞) ) ≤ + M δ Rη − ln θ

(4.127)

From this, we obtain that cη ln(2R) − ln δ ln Nδ (G (∞) ) η(ln(2R) − ln δ ) 1 ˜0 ≤ + ln + M − ln δ − ln δ − ln δ Rη − ln θ =: h3 (δ ) . We conclude then that if θ η H1 (θ ) ≤ 1, dimF (G (∞) ) ≤ lim h3 (δ ) = η . δ →0

(4.128)

If instead θ η H1 (θ ) ≥ 1, from (4.116), (4.118) and (4.124) we have η cη 2R Nδ (G (∞) ) ≤ η (θ η H1 (θ ))n+1 (n + 1) + M˜ 0 (H1 (θ ))n+1 R δ η ln(2R/δ ) ln(2R/δ ) ln(2R/δ ) cη 2R ≤ η + M˜ 0 (H1 (θ )) − ln θ (θ η H1 (θ )) − ln θ R δ − ln θ ln(H (θ )) 1 1 (θ )) η − ln θ −η ln(H − ln θ cη 2R ln(2R/δ ) 2R 2R = η + M˜ 0 R δ δ − ln θ δ Θ cη ln(2R) − ln δ 2R ≤ + M˜ 0 . δ Rη − ln θ Since Θ ≤ η, this yields the same estimate as (4.127); hence, we conclude that (4.128) holds also if θ η H1 (θ ) ≥ 1. Thus, (4.115) holds, and this concludes the proof of (4.114). 8d. We can now conclude the proof of the estimate of the fractal dimension of the exponential attractor E. In fact, recalling (4.83), (4.15) follows from (4.87) and (4.114), using the first of part (3) of proposition 2.61. With this, we have finally completed the proof of theorem 4.26: the set E defined in (4.83) is the desired exponential attractor of the semiflow S, relative to B. Finally, since theorem 4.4 follows from theorem 4.26, theorem 4.4 is now also completely proven.

4.6

Concluding Remarks

175

4.6 Concluding Remarks We conclude this chapter by remarking that the construction of the exponential attractor that we have described is not the only possible one. In fact, in Eden, Foias, Nicolaenko and Temam, [EFNT94, ch. 7], an alternative construction of the exponential attractor is given, which is not based on the cone property. We refer to [EFNT94] for all details; here, we limit ourselves to illustrate one of these alternative ways to obtain an exponential attractor, by considering the following example, which is adapted from [EFNT94, sct. 7.1]. Example 4.27 Let K := {x ∈ RN : kxk ≤ 1} be the closed unit ball of RN , and define a map S : K → K by x S(x) := . 1 + kxk Then, it is easily seen by induction that x , Sn (x) = 1 + nkxk for all n ∈ N and x ∈ K. Thus, S has the attractor A = {0}, but the convergence of Sn (x) to A is only polynomial. In fact, note that for all x ∈ K and n ≥ 1, kSn (x)k ≤ while for all x ∈ K \ {0} and n ≥

j

1 kx k

k

1 , n

+ 1,

kSn (x)k ≥

1 . 2n

To construct an exponential attractor E, we proceed in the following way. We start by choosing, as a compact absorbing set B, any closed ball B(0, R), 0 < R ≤ 1. Thus, we look for points in K which attract other points exponentially, without necessarily converging exponentially to 0. Given θ ∈ ]0, 1[, we consider the set G := {z ∈ K : z = kθ m x, x ∈ K, k, m ∈ N, kθ m ≤ 1} . The set G is evidently compact. We claim that G attracts all subsets D ⊆ K exponentially. Indeed, given any D ⊆ K and n ∈ N, for each x ∈ D we define h ∈ ]0, +∞[ and k ∈ N by 1 , k := bhc h := (1 + nkxk)θ n (both h and k depend on x and n). Since k ≤ h, kθ n ≤ hθ n =

1 ≤ 1, 1 + nkxk

176

4

Exponential Attractors

which implies that z := kθ n x ∈ G. Since also h < k + 1, and kxk ≤ 1, kSn (x) − zk ≤ kSn (x) − hθ n xk + khθ n x − kθ n xk = 0 + (h − k)θ n kxk ≤ θ n . Thus, for each x ∈ D, d(Sn (x), G) = inf d(Sn (x), z) ≤ kSn (x) − zk ≤ θ n . z∈G

Since this estimate is independent of x ∈ D, and in fact of D, we conclude that for all D ⊆ K, ∂ (Sn (D), G) ≤ θ n = e−κn , with κ := − ln θ . However, the set G is not yet an exponential attractor, since it needs not be positively invariant. Thus we enlarge it into the set E=

∞ [

Sn (G) .

n=0

This set is now positively invariant, again compact, and attracts all subsets of K exponentially. The fractal dimension of E is obviously finite (being at most equal to N); hence, E is the desired exponential attractor.

Chapter 5 Inertial Manifolds

5.1 Introduction 1. Roughly speaking, INERTIAL MANIFOLDS are positively invariant, finite dimensional, exponentially attracting Lipschitz manifolds. In this chapter we give the precise definition of this notion for a continuous semiflow S defined on a Banach space X , and show how an inertial manifold can be constructed, when S satisfies some natural conditions on its geometrical structure. In this chapter, we study in detail inertial manifolds which are graphs of maps m : X1 → X , with X1 a finite dimensional subspace of X . These manifolds have a degree of smoothness, inherited by the smoothness of the map m (we shall in particular consider Lipschitz continuous maps), and the geometrical structure of the semiflow S can be described quite naturally for manifolds of this type. In particular, we shall consider the CONE INVARIANCE PROPERTY, and one of several versions of the STRONG SQUEEZING PROPERTY. We proceed then to explore the applicability of these results to abstract evolution equations of the form ut + A u = F(u) ,

(5.1)

where A : X → X is a linear, unbounded operator which generates at least a C0 SEMIGROUP in X (see section 5.6.1 below). We find that the squeezing property holds if the eigenvalues in the point spectrum of A satisfy some restrictions on their growth, relative to the nonlinearity F; in particular, if they satisfy one type of SPEC TRAL GAP CONDITION . For problems of the type considered in chapter 3, we will be able to show that, when the space dimension is n = 1, this spectral gap condition always holds for the parabolic problem (P), as well as for the hyperbolic problem (Hε ), if ε is sufficiently small. For higher space dimensions, the situation depends heavily on special features of the problem, such as the geometric properties of the domain Ω . In particular, we will consider in detail a model of the C HAFEE -I NFANTE reaction-diffusion equations in one dimension of space, and its (small) hyperbolic perturbation. In chapter 6 we present some other examples of semiflows which admit an inertial manifold; mostly, these semiflows are generated by PDEEs of “parabolic” type. 2. The limitations on the possibility of extending this theory to general dissipative hyperbolic problems are highlighted by a remarkable result of Mora and SolàMorales ([MSM87]), concerning a one-dimensional version of the dissipative wave

177

178

5

Inertial Manifolds

equation (3.4). In short, they have shown that if ε is sufficiently large, and the boundary conditions for (3.4) are of Neumann type, the corresponding semiflow does not admit any inertial manifold which is of class C1 and locally invariant in the neighborhood of one of the stationary solutions of the problem. Thus, for general ε > 0 the existence of an inertial manifold in the hyperbolic case is open to question. We present Mora and Solà-Morales result in chapter 7. 3. When the global attractor A and a closed inertial manifold M exist, then A ⊆ M. This follows in a similar way as the inclusion of the global attractor A in an exponential attractor E (see section 4.1). On the other hand, if G ⊆ X is a bounded, positively invariant absorbing set, theorems 2.33 and 2.46 imply that the semiflow S admits a global attractor A ⊆ G, if either G is compact or if S is uniformly compact for large t. In either case, if S admits both a compact, positively invariant absorbing set G, and a closed inertial manifold M, then the set E := M∩G is also a compact set, which is positively invariant (being the intersection of two positively invariant sets). If M possesses a more specific type of attractivity property, called EXPONENTIAL TRACKING PROPERTY , then E is exponentially attracting (see section 5.2.2). Thus, in this case, E is an exponential attractor, and A ⊆ E ⊆ M. 4. As we have already mentioned, there are many more systems that admit an exponential attractor than systems that are known to admit an inertial manifold. The main reason for this difference is that all known inertial manifolds are closed, and therefore the existence of a compact absorbing set (which does hold for the systems we have considered so far) also yields directly the existence of an exponential attractor. Moreover, inertial manifolds are much more regular than exponential attractors. There is also a “practical” reason for this difference; namely, that the available results on the existence of inertial manifolds require two conditions on the geometrical structure of the semiflow, called here respectively the CONE INVARIANCE PROPERTY and the DECAY PROPERTY. These are usually combined together, in the so-called STRONG SQUEEZING PROPERTY . On the other hand, in order to establish the existence of an exponential attractor it is sufficient to assume that the semiflow S satisfies the discrete squeezing property, which is a much weaker condition. 5. Roughly speaking, the cone invariance and decay properties describe a sort of dichotomy principle, whereby either the difference of two motions can never leave a certain cone (cone invariance property), or, if it does, the distance between the motions decays exponentially (decay property). In contrast, the discrete squeezing property only requires that either the difference of two motions is in a cone at a specific time (as opposed to for all times), or, if not, the distance between the motions decays exponentially (as in the decay property). For semiflows generated by evolution equations of the general form (5.1), one of the available ways to verify that the semiflow satisfies the cone invariance property is to deduce this property from a condition on the point spectrum of A. This condition

5.2

Definitions and Comparisons

179

is called the SPECTRAL GAP CONDITION; we introduce it in section 5.6.2. Essentially, this condition guarantees that the linear part of the equation (i.e., the term Au) can “dominate”, in a sense to be made precise, the nonlinear part (i.e., the term f (u)). In most situations of interest, the spectral gap condition is quite difficult if not impossible to verify, while the decay properties are relatively easy to verify. This is not surprising, since the spectral gap condition requires a sufficiently large difference between successive eigenvalues, whereas, for the decay property, one only needs a sufficiently large eigenvalue. 6. As we have stated, we will construct inertial manifolds of the type M = {x + m(x) : x ∈ X1 } ,

(5.2)

where X = X1 ⊕ X2 is decomposed into a closed linear subspace X1 of finite dimension, and its algebraic complement X2 , and where m : X1 → X2 is a Lipschitz continuous map. In contrast to exponential attractors, this type of inertial manifolds allows us to imbed the global attractor A (and the exponential attractor E) in RN , with N = dim X1 = dim M. Moreover, if A commutes with the continuous projector π1 from X onto X1 , the asymptotic behavior of the solution of (4.4) is governed by the N-dimensional INER TIAL FORM SYSTEM

x˙ = −Ax + π1 f (x + m(x))

(5.3)

in X1 . This is the so-called SLAVING PRINCIPLE. In contrast to (4.3), (5.3) is explicitly given and has a Lipschitz continuous right-hand side if f is Lipschitz continuous.

5.2 Definitions and Comparisons In this section we give the definition of an inertial manifold, and compare various ways that can be used to construct an inertial manifold.

5.2.1 Lipschitz Manifolds and Inertial Manifolds In the literature on inertial manifolds, the notion of Lipschitz manifold is often not defined precisely. In most applications, there are two types of manifolds M which are called Lipschitz manifolds: In the first, the manifold M is the graph of a Lipschitz continuous map m from a closed, linear subspace X1 of X into its algebraic complement X2 , as in (5.2). In the second, M is the graph of a Lipschitz continuous map m from a bounded subset M1 of X1 into X2 , where the set M1 can be closed. In both cases, X1 is typically finite dimensional; at least for the second case, we need the notion of a Lipschitz manifold with boundary.

180

5

Inertial Manifolds

We denote by HN := {(x1 , . . . , xN ) ∈ RN : x1 ≥ 0} the (first) closed half-space of RN . DEFINITION 5.1 A subset M ⊆ X is an N- DIMENSIONAL L IPSCHITZ SUBMAN IFOLD ( WITH BOUNDARY ) of X if it possesses the following two structural characteristics: 1. There exists a countable S collection of open sets Vi ⊆ X , i ∈ I, such that if Ui := Vi ∩ M, then M = i∈I Ui . 2. There exist open sets Wi ⊆ RN and invertible mappings Φi : Wi → Vi , with Φi (Wi ∩ HN ) = Ui , such that Φi and Φi−1 are Lipschitz continuous. The set ∂ M :=

[

Φi (∂ HN )

i∈I

is called the BOUNDARY of M. In particular, we have the so-called trivial Lipschitz submanifolds: DEFINITION 5.2 A subset M ⊆ X is called an N- DIMENSIONAL TRIVIAL L IP SCHITZ SUBMANIFOLD ( WITH BOUNDARY ) of X if it possesses the following two structural characteristics: 1. There exist a closed N-dimensional linear subspace X1 of X , and a subset M1 of X1 , such that M1 is an N-dimensional Lipschitz manifold (with boundary). 2. There is an invertible mapping ϕ : M1 → M such that ϕ and ϕ −1 are Lipschitz continuous. REMARK 5.3 1. A trivial N-dimensional Lipschitz submanifold (with boundary) of X is indeed an N-dimensional Lipschitz submanifold of X . To show this, in accord with definition 5.2, let Vi ⊆ X , Wi ⊆ RN be open sets and let Φi : Wi ∩ HN → Ui , S Ui := Vi ∩ M1 , be invertible mappings such that M1 = i∈I Ui and such that Φi and Φi−1 are Lipschitz continuous. Let V˜ i := ϕ(Vi ). Then the sets V˜ i are open sets, S as pre-images of the open sets Vi , and M = i∈I U˜i with U˜i := V˜ i ∩ M. Finally, Φ˜ i := ϕ ◦ Φi maps Wi ∩ HN onto U˜i and Φ˜ i and Φ˜ i−1 are Lipschitz continuous. 2. If, in addition, M1 is closed, then M is also a closed set. This follows from the fact that M is the pre-image of the closed set M1 under the continuous mapping ϕ −1 . 3. We will consider only two particular cases: The first is the global one, in which M1 = X1 . The second is when M1 is a bounded, closed subset of X1 , with Lipschitz boundary.

5.2

Definitions and Comparisons

181

Example 5.4 1. The set M := {(x, y, z) ∈ R3 : max{|x|, |y|} ≤ 1, z = xy} is a two dimensional, trivial Lipschitz submanifold of R3 , with boundary. The boundary ∂ M of M is the set {(x, y, z) ∈ R3 : z = xy, |x| = 1 or |y| = 1} . 2. The set M := {(x, y, z) ∈ R3 : max{|x|, |y|, |z|} = 1, z ≥ 0} is a two dimensional, nontrivial Lipschitz submanifold of R3 , with boundary given by the set {(x, y, z) ∈ R3 : z = 0, |x| = |y| = 1} .

We can now introduce the definition of inertial manifold. DEFINITION 5.5 Let T be one of the sets R or R≥0 , and let S = (S(t))t ∈T be a continuous semiflow on X . A subset M ⊂ X is an INERTIAL MANIFOLD for S if M is a finite dimensional Lipschitz submanifold (with boundary) of X , which is positively invariant and exponentially attracting. The latter condition means that there is η > 0 with the property that for all x ∈ X there is K > 0 such that for all t ≥ 0, d(S(t)x, M) ≤ K e−ηt .

(5.4)

REMARK 5.6 1. In the naive understanding of Lipschitz manifolds, the difference between manifolds with or without boundary is not usually emphasized. 2. Currently, we are only able to construct trivial Lipschitz manifolds as graphs of a Lipschitz mapping m from a (closed) subset M1 of a (closed) linear subspace X1 of X into the algebraic complement X2 of X1 , as in (5.2). 3. Inertial manifolds which are trivial Lipschitz manifolds, in the sense of definition 5.2, can be assumed to be closed (at least after extending of m to the closure of its domain.) The notion of inertial manifolds for evolution equations goes back to at least Henry, [Hen81], and Mora, [Mor87], and was also studied by Constantin, Foias, Nicoalenko, Sell and Temam, [FST85, FNST85, CFNT86]; see also Temam, [Tem88]. In all these works, the inertial manifolds were constructed as the intersection of a graph of Lipschitz function over a closed, linear, N-dimensional subspace X1 of X , with a closed ball in X . Thus, these manifolds are compact, and possess a finite fractal dimension (which is N). Together with the exponential convergence of the orbits described in (5.4), this is a property shared with exponential attractors; as we discussed in chapter 4, these properties imply that, when an inertial manifold exists, this set

182

5

Inertial Manifolds

A

E M

Figure 5.1: Inertial Manifold, Absorbing Set, and Global Attractor provides an extremely effective way of describing the long-time behavior of the dynamical system, since the latter is asymptotically equivalent to a finite dimensional system (that is, to a system of ODEs). One of the advantages of inertial manifolds over exponential attractors resides in the fact that, when an inertial manifold exists, this system of ODEs has smooth coefficients, inherited from the smoothness of M. In many situations, it is convenient to strengthen the requirement that motions converge exponentially to M, as in (5.4), and assume the following property: DEFINITION 5.7 Let M ⊆ X be a finite-dimensional Lipschitz manifold, positively invariant with respect to S. M is said to have the EXPONENTIAL TRACKING PROPERTY , if there is η > 0 such that for every x ∈ X , there are x0 ∈ M and c ≥ 0 such that for all t ≥ 0, kS(t)x − S(t)x0 kX ≤ ce−ηt

(5.5)

(note that S(t)x0 ∈ M for all t ≥ 0, because M is positively invariant.) The motion t → S(t)x0 in (5.5) is called an ASYMPTOTIC PHASE of the motion t → S(t)x. The exponential tracking property, which is also called EXISTENCE OF ASYMP or ASYMPTOTIC COMPLETENESS PROPERTY, was introduced, for example, by Foias, Sell and Titi in [FST89], and by Robinson in [Rob96]. It obviously implies the exponential convergence of the motions to the manifold, because if x ∈ X and x0 ∈ M are as in (5.5), then, since S(t)x0 ∈ M for all t ≥ 0, TOTIC PHASES ,

d(S(t)x, M) = inf kS(t)x − ykX ≤ kS(t)x − S(t)x0 kX ≤ ce−ηt , y∈M

(5.6)

i.e., (5.4) holds. Of course, the exponential tracking property is much stronger than the exponential convergence of the motions, since (5.5) means that any motion should converge exponentially to a motion which is completely in the manifold M.

5.2

Definitions and Comparisons

183

In (5.5), the constant c depends on x, both explicitly and via x0 , which also depends on x. In practice, it is often possible to show that even stronger versions of the exponential tracking property hold. For example, we can require exponential tracking properties with inequalities of the form kS(t)x − S(t)x0 kX ≤ K1 kx − x0 kX e−ηt , kS(t)x − S(t)x0 kX ≤ K2 d(x, M)e−ηt ,

(5.7)

with K1 and K2 independent of x ∈ X and x0 ∈ M. For example, (5.7) means that the constant c in (5.5) does not depend on the point x0 ∈ M, but only on the distance of x to the manifold M. In other words, the motion t 7→ S(t)x admits the motion t 7→ S(t)x0 as an asymptotic phase, and the exponential decay of the difference between these motions is estimated by the initial distance between x and the manifold M, with constants K2 and η independent of x. Finally, there is no loss in generality in considering x ∈ X \ M only, because if x ∈ M we can take x0 = x, and (5.7) is trivially satisfied.

5.2.2 Inertial Manifolds and Exponential Attractors In general, the existence of an inertial manifold does not imply, and is neither implied by, the existence of an exponential attractor. Moreover, even when both these sets exist, neither is in general contained in the other (although they both contain the global attractor). However, we can construct an exponential attractor from an inertial manifold, by intersecting the latter with a compact, positively invariant absorbing set. More precisely: PROPOSITION 5.8 Let S be a semiflow on X . Assume that S admits a compact, positively invariant absorbing set G, and a closed, inertial manifold M. Assume further that the stronger version (5.7) of the exponential tracking property holds. Then the set E := M ∩ G is an exponential attractor for S in X . PROOF Since M is closed and G is compact, E is compact. Since both M and G are positively invariant, E is positively invariant too. Since E ⊆ M, by (2.49) of proposition 2.61 we have dimF (E) ≤ dimF (M), so E has finite fractal dimension. Thus, it only remains to prove that E attracts all bounded sets of X exponentially, i.e., that for any bounded set B ⊆ X , there is CB > 0, depending on B, such that for all t ≥ 0, ∂ (S(t)B, E) ≤ CB e−ηt .

(5.8)

As an intermediate step, we show that (5.8) holds when B = G. To this end, we turn to the modified exponential tracking property (5.7). Since the function x 7→ d(x, M) is continuous, its restriction to the compact set G is bounded; therefore, by (5.7), there

184

5

Inertial Manifolds

is KG > 0, depending on G, such that for all g ∈ G, there is g0 ∈ M with the property that for all t ≥ 0, kS(t)g − S(t)g0 kX ≤ KG e−η t .

(5.9)

Call G 0 the subset of M consisting of all the elements g0 ∈ M satisfying (5.9) for some g ∈ G. Then, G 0 is bounded. In fact, from (5.9) for t = 0 we have kg0 kX ≤ kg0 − gkX + kgkX ≤ KG + sup kxkX . x∈G

Since G is absorbing, there is T1 ≥ 0, depending on G 0 , such that for all t ≥ T1 , S(t)G 0 ⊆ G. Since M is positively invariant, we deduce that if t ≥ T1 , S(t)G 0 ⊆ G ∩ M = E . Hence, (5.9) implies that for all g ∈ G and t ≥ T1 , d(S(t)g, E) ≤ kS(t)g − S(t)g0 kX ≤ KG e−ηt .

(5.10)

Since the right side of (5.10) is independent of g, it follows that, if t ≥ T1 , ∂ (S(t)G, E) ≤ KG e−ηt .

(5.11)

If instead 0 ≤ t ≤ T1 , we estimate ∂ (S(t)G, E) ≤ max ∂ (S(t)G, E) =: δ1 = δ1 eηt e−ηt ≤ δ1 eηT1 e−ηt . 0≤t ≤T1

(5.12)

Thus, when B = G, (5.8) follows from (5.12) and (5.11), with CG := max{KG , δ1 eηT1 } . We now show (5.8) for a general bounded set B ⊆ X . Since G is absorbing, there is T2 ≥ 0, depending on B, such that for all t ≥ T2 , S(t)B ⊆ G. Let t ≥ T2 , and write t = T2 + θ , θ ≥ 0. Then, since S(t)B = S(θ )S(T2 )B ⊆ S(θ )G , we have that when t ≥ T2 , ∂ (S(t)B, E) ≤ ∂ (S(θ )G, E) ≤ CG e−ηθ = CG e−η(t −T2 ) = CG eηT2 e−ηt .

(5.13)

If instead 0 ≤ t ≤ T2 , we proceed as in (5.12), i.e. ∂ (S(t)B, E) ≤ max ∂ (S(t)B, E) =: δ2 ≤ δ2 eηT2 e−ηt . 0≤t ≤T2

Thus, (5.8) follows from (5.13) and (5.14), with CB := eηT2 max{CG , δ2 } . This concludes the proof of proposition 5.8.

(5.14)

5.2

185

Definitions and Comparisons

5.2.3 Methods of Construction of the Inertial Manifold 1. Most available ways to construct an inertial manifold are based on generalizations of methods developed for the construction of unstable, center-unstable or center manifolds for ODEs. For an extensive review, we refer e.g. to Luskin and Sell, [LS89], or to Ninomiya, [Nin93]. Here, we briefly recall the so-called LYAPUNOVP ERRON , the H ADAMARD and the integral manifold methods. The Lyapunov-Perron method was introduced by Lyapunov, [Lya47, Lya92] and Perron, [Per28, Per29, Per30], to prove the existence of stable and unstable manifolds of hyperbolic equilibrium points of systems of ODEs. In this method, the ODEs are transformed into integral equations, and the invariant manifolds are constructed as fixed points of the corresponding integral operator. In [Hen81], Henry gives a generalization of the Lyapunov-Perron method, which leads to a proof of the existence of stable, unstable and center manifolds for semilinear parabolic evolution equations. Further generalizations to the construction of inertial manifolds are presented e.g. in Foias, Sell and Temam, [FST88], Temam, [Tem88], Constantin, Foias, Nicolaenko and Temam, [CFNT86], Foias, Sell and Titi, [FST89], and Demengel and Ghidaglia, [DG91]. Indeed, the Lyapunov-Perron method is arguably one of the most commonly used for the construction of inertial manifolds. Hadamard’s method, introduced in [Had01] and also known as the GRAPH TRANS FORMATION METHOD , has been developed to show the existence of stable and unstable manifolds of fixed points of diffeomorphisms. Its nature is more geometrical than the Lyapunov-Perron method, in that the stable and unstable manifolds (relative to hyperbolic fixed points) are constructed as graphs over the linearized stable and unstable subspaces. An extension of Hadamard’s method to infinite dimensional dynamical systems can be found e.g. in Bates and Jones, [BJ89]. The integral manifold method was introduced by Constantin, Foias, Nicoalenko and Temam in [CFNT89]. In this construction, the inertial manifold is defined as the S closure of the set t ≥0 S(t)Γ , where Γ is the C1 boundary of a suitable closed subset of a finite-dimensional, closed, linear subspace of X . 2. In all these three methods, the inertial manifold M is constructed as the graph of a Lipschitz continuous function m defined on a subset M1 of a finite dimensional, closed, linear subspace X1 of X . More precisely, we seek to establish a suitable orthogonal decomposition X = X1 ⊕ X2 ,

(5.15)

where X1 is a finite dimensional subspace of X and the orthogonal projections πi : X → Xi ,

i = 1, 2

are continuous, and to construct an inertial manifold M for the semiflow S, having the form M = graph(m) := {ξ + m(ξ ) : ξ ∈ M1 } ,

(5.16)

186

5

Inertial Manifolds

where m : M1 ⊆ X1 → X2 is a Lipschitz continuous function. At least if M1 = X1 , M is a Lipschitz manifold; see figure 5.2. Indeed, let M1 := X1 , and define X2 M X1

Figure 5.2: A global trivial Lipschitz manifold M = graph m ϕ : M1 → X by ϕ(x) := x + m(x), x ∈ X1 . Obviously, ϕ and ϕ −1 = π1 are Lipschitz continuous. If M1 6= X1 , in order for M to be a Lipschitz manifold we need M1 to be a Ndimensional Lipschitz submanifold of the N-dimensional space X1 . For example, if M1 is a closed ball in X1 , if m˜ : X1 → X2 is Lipschitz continuous, and if m := m˜ M 1 is the restriction of m˜ onto M1 , then M is a Lipschitz manifold with boundary; see figure 5.3. Indeed, the maps ϕ : M1 → M and ϕ −1 : M → M1 defined by X2 X1

M

C Figure 5.3: A trivial Lipschitz manifold M = (graph m) ∩ C as intersection of a global trivial Lipschitz manifold with a cylinder C

ϕ(x) := x + m(x) ,

x ∈ M1 ,

ϕ −1 := π1 C ,

where C := {x ∈ X : π1 x ∈ M1 } is the cylinder in X with base M1 , are Lipschitz continuous. 3. Here, we shall follow Hadamard’s method; that is, we obtain the function m as a fixed point of a so-called GRAPH TRANSFORMATION MAP, defined on the subspace

5.2

Definitions and Comparisons

187

GL of the Banach space G := Cb (X1 ; X2 ) ,

(5.17)

consisting of all continuous and bounded functions from X1 into X2 , which satisfy a global Lipschitz condition of the form km(ξ ) − m(η)kX ≤ L kξ − ηkX ,

ξ , η ∈ X1 ,

(5.18)

where X1 is as in (5.15) (we require the boundedness of the functions, in order to make GL a proper Banach space of continuous functions). That is, given L > 0 we define GL := {ψ ∈ G : kψ(ξ ) − ψ(η)kX ≤ Lkξ − ηkX for all ξ , η ∈ X1 } ,

(5.19)

and we consider the constant L as a parameter for the set GL , in which we will eventually find the function m, whose graph will be the desired inertial manifold. 4. As noted above, M is closed, being the pre-image of the closed set X1 under the Lipschitz map π1 . It follows that if the semiflow S admits a global attractor A and an inertial manifold M of the form (5.16), then A ⊆ M. On the other hand, inertial manifolds of the form (5.16) need not be compact, since they are not necessarily bounded. To obtain inertial manifolds that are compact, it is sufficient to intersect the inertial manifold (5.16) with a compact, positively invariant absorbing set. More precisely: PROPOSITION 5.9 Let S be a continuous semiflow on X , and assume that there is a Lipschitz continuous function m as in (5.16), whose graph M is an inertial manifold for S. Assume that M satisfies the exponential tracking property (5.7), and that S admits a compact, ˜ := M ∩ B is a nonempty set with Lipschitz positively invariant absorbing set B. If M ˜ is a compact inertial manifold for S (and also an exponential boundary, then M attractor). ˜ ⊆ B, and B is compact, M ˜ is also compact. ExPROOF Since M is closed, M ˜ actly as in the proof of proposition 5.8, we see that M is positively invariant and ˜ has finite fractal dimension, we note that exponentially attracting. To show that M ˜ M is the image of the compact set π1 (B) under the Lipschitz continuous map m. Hence, by proposition 2.61, ˜ ≤ dimF (π1 (B)) ≤ dimF (X1 ) . dimF (M) ˜ is an inertial manifold for S. Because of proposition This completes the proof that M ˜ is also an exponential attractor. 5.8, M

188

5

Inertial Manifolds

For future reference, we recall here that, since the projections πi are linear, their continuity is equivalent to the boundedness conditions kπi kL(X ,X ) := sup kπi xkX < +∞ ,

i = 1, 2 .

(5.20)

kxkX ≤1

Actually, since π1 + π2 = IX (the identity in X ), it is sufficient to assume (5.20) for i = 1 or i = 2 only. 5. Finally, we would like to comment on the significance of the exponential attracting property (5.4), in relation to manifolds of the form (5.16). As we have stated above, if S admits an inertial manifold M, then motions on M are essentially governed by a finite system of ODEs. These motions are such that S(t)x ∈ M for all t ≥ 0; if M has the form (5.16), this condition translates into the identity π2 S(t)x = m(π1 S(t)x) . When instead the motion does not take place on M, i.e. when S(t)x ∈ / M, we want the difference R(t)x := π2 S(t)x − m(π1 S(t)x) , which does no longer vanish, to become negligible as t → +∞. More specifically, we require this difference to decay exponentially, i.e. kR(t)xkX ≤ Ce−ηt ,

(5.21)

for some positive C and η, the latter independent of x. Now, since π1 S(t)x + m(π1 S(t)x) ∈ M ,

(5.22)

we have that, for all z = ξ + m(ξ ) ∈ M, kR(t)xkX ≤ kπ2 S(t)x − m(ξ )kX + km(ξ ) − m(π1 S(t)x)kX ≤ kπ2 S(t)x − π2 zkX + Lkξ − π1 S(t)xkX = kπ2 (S(t)x − z)kX + Lkπ1 (z − S(t)x)kX ≤ (1 + L)kS(t)x − zkX (recall that L denotes the Lipschitz constant of m). Since z is arbitrary in M, it follows that kR(t)xkX ≤ (1 + L)d(S(t)x, M) . Thus, (5.21) does hold if M is exponentially attracting, i.e. if (5.4) holds. Incidentally, we also note that, conversely, (5.4) follows from (5.21). Indeed, (5.22) also implies that d(S(t)x, M) ≤ kS(t)x − (π1 S(t)x + m(π1 S(t)x))kX

5.3

Geometric Assumptions on the Semiflow

189

= k(S(t)x − π1 S(t)x) − m(π1 S(t)x)kX = kπ2 S(t)x − m(π1 S(t)x)kX = kR(t)xkX . In the same way, we see that, for example, the exponential tracking property (5.7) implies the estimate kπ2 S(t)x − m(π1 S(t)x)kX ≤ Cd(x, M)e−ηt ,

(5.23)

for all x ∈ X and t ≥ 0.

5.3 Geometric Assumptions on the Semiflow In this section we introduce two geometric assumptions on the semiflow S, that we first recognize as “natural” conditions that S should satisfy in order to admit an inertial manifold of the form (5.16), and then show to be almost sufficient for the construction of such type of an inertial manifold. These properties are the CONE IN VARIANCE PROPERTY and the STRONG SQUEEZING PROPERTY . To describe these conditions, it is useful to keep in mind that our goal is the determination of a finite dimensional subspace X1 of X , and of a function m ∈ GL , such that the graph of m is an inertial manifold for S, as in (5.16).

5.3.1 The Cone Invariance Property We first introduce the CONE INVARIANCE PROPERTY. Given L > 0, we denote by CL the cone CL := {x ∈ X : kπ2 xkX ≤ Lkπ1 xkX } .

(5.24)

We immediately have PROPOSITION 5.10 Let G and GL be as in, respectively, (5.17) and (5.19). Let m ∈ G. Then m ∈ GL if and only if for all x, y ∈ graph(m), x − y ∈ CL . PROOF The proof is immediate. Assume first that m ∈ GL . Recalling (5.24), we need to prove that if x, y ∈ graph(m), kπ2 (y − x)kX ≤ Lkπ1 (y − x)kX .

(5.25)

But, by (5.16), x = ξ + m(ξ ) and y = η + m(η) for some ξ and η ∈ π1 X , with m(ξ ) and m(η) ∈ π2 X . Consequently, π2 (y − x) = m(η) − m(ξ ) and π1 (y − x) = η − ξ ,

190

5

Inertial Manifolds

and (5.18) implies (5.25). Conversely, let m ∈ G, and assume that for all x, y ∈ graph(m), y − x ∈ CL . We show that m satisfies (5.18). Thus, given ξ and η ∈ π1 X , let x = ξ + m(ξ ) and y = η + m(η). Then x and y ∈ graph(m), so y − x ∈ CL , and we easily conclude as in the first part of the proof. COROLLARY 5.11 Let m ∈ GL , and M := graph(m). If M is positively invariant, then for all x1 , x2 ∈ M, and all t ≥ 0, S(t)x1 − S(t)x2 ∈ CL . PROOF Let x1 , x2 ∈ M. Since M is positively invariant, S(t)x1 and S(t)x2 ∈ M for all t ≥ 0, and S(t)x1 − S(t)x2 ∈ CL , by proposition 5.10. From proposition 5.10 and corollary 5.11 we deduce that if the semiflow S did have an inertial manifold M of the form (5.16), that is, if there were m ∈ GL such that M = graph(m), then for all x1 and x2 ∈ M, both x1 −x2 ∈ CL and S(t)x1 −S(t)x2 ∈ CL for all t ≥ 0 (the latter because of the positive invariance of M). We now note that these two conditions can be formulated independently of the actual knowledge of M (which, of course, is the very manifold we want to construct). It is therefore natural to require this proposition as an a priori condition on the semiflow S. Thus, we replace the assumption x1 , x2 ∈ M (which we cannot state if M is still to be found), with the assumption x1 − x2 ∈ CL (which we can state), and require that this new assumption implies that the corresponding difference S(t)x1 − S(t)x2 be in CL for all t ≥ 0, as was the case in corollary 5.11. This motivates the following DEFINITION 5.12 Let L > 0. The continuous semiflow S satisfies the CONE VARIANCE PROPERTY with parameter L if for all x1 , x2 ∈ X , and all t ≥ 0, x1 − x2 ∈ CL =⇒ S(t)x1 − S(t)x2 ∈ CL .

IN -

(5.26)

Once again, the cone invariance property means that if the difference x1 − x2 of two points is in the cone CL , then the difference S(t)x1 − S(t)x2 of all successive points on the corresponding orbits starting at x1 and x2 remains in the same cone CL for all future times. That is, the cone CL is invariant with respect to the difference of forward motions (see figure 5.4). The following result is a direct consequence of the cone invariance property (5.26): PROPOSITION 5.13 Assume the continuous semiflow S satisfies the cone invariance property with parameter L. Let x1 , x2 ∈ X , and assume there exists T > 0 such that S(T )x1 − S(T )x2 6∈ CL . Then S(t)x1 − S(t)x2 6∈ CL also for all t ∈ [0, T ]. PROOF We proceed by contradiction. If there were t 0 ∈ [0, T ] such that S(t 0 )x1 − S(t 0 )x2 ∈ CL , then, setting t = T −t 0 ≥ 0 in (5.26) with xi replaced by S(t 0 )xi , i = 1, 2,

5.3

Geometric Assumptions on the Semiflow

191

S(t)x1 + CL x1 + CL S(t)x1

x1 x2

S(t)x2

Figure 5.4: The Cone Invariance Property we would obtain that S(t)S(t 0 )x1 − S(t)S(t 0 )x2 ∈ CL . But S(t)S(t 0 ) = S(T ), so this cannot hold.

5.3.2 Decay and Squeezing Properties In this section we introduce the squeezing properties as another set of natural conditions that the continuous semiflow S should satisfy in order to admit an inertial manifold M of the form (5.16), which also satisfies the modified version (5.7) of the exponential tracking property. To derive these additional conditions, we again start by assuming that S does have an inertial manifold with the desired properties. In addition, we assume that S satisfies the cone invariance property (5.26), and that it admits a positively invariant manifold M = graph(m), with m ∈ GL0 ⊂ GL for some L0 < L (that is, the function m satisfies a stronger Lipschitz condition). This additional assumption is not so restrictive, since in most applications it often happens that the parameter L is determined as a solution of an inequality of the form F(L) > 0, where F : ]0, ∞[ → R is a continuous function. Therefore, since we can take L0 < L sufficiently close to L and still have F(L0 ) > 0, in this case the semiflow does satisfy a second cone invariance property, as stated. (An instance when such a second cone invariance property is explicitly used can be found in Robinson, [Rob93, prop. 3].) Under these conditions, we claim: PROPOSITION 5.14 There is a constant c2 > 0 such that for all x, y and z ∈ X , if x − z 6∈ CL but y − z ∈ CL0 , kx − zkX ≤ c2 kx − ykX .

(5.27)

PROOF Since x − z 6∈ CL , (5.24) implies that Lkπ1 (x − z)kX < kπ2 (x − z)kX ; on the other hand, since y − z ∈ CL0 , again (5.24) implies that kπ2 (y − z)kX ≤ L0 kπ1 (y − z)kX .

(5.28)

192

5

Inertial Manifolds

Consequently, we estimate Lkπ1 (y − z)kX ≤ Lkπ1 (y − x)kX + Lkπ1 (x − z)kX ≤ Lkπ1 (y − x)kX + kπ2 (x − z)kX ≤ Lkπ1 (y − x)kX + kπ2 (x − y)kX + kπ2 (y − z)kX ≤ Lkπ1 (y − x)kX + kπ2 (x − y)kX + L0 kπ1 (y − z)kX . Since L > L0 , we obtain (L − L0 )kπ1 (y − z)kX ≤ Lkπ1 (y − x)kX + kπ2 (x − y)kX ≤ (1 + L)kx − ykX . Consequently, kx − zkX ≤ kx − ykX + ky − zkX ≤ kx − ykX + kπ1 (y − z)kX + kπ2 (y − z)k ≤ kx − ykX + (1 + L0 ) kπ1 (y − z)kX ≤ kx − ykX +

(1 + L0 )(1 + L) kx − ykX , L − L0

and (5.27) follows. Let now x1 ∈ X \ M, and assume there are x2 ∈ M and t > 0 such that S(t)x1 − S(t)x2 ∈ / CL .

(5.29)

Let w := π1 S(t)x1 + m(π1 S(t)x1 ). Then, w ∈ M. Since also S(t)x2 ∈ M (because M is positively invariant), by proposition 5.10 it follows that S(t)x2 − w ∈ CL0 . Thus, by (5.29), we can choose x = S(t)x1 , y = w and z = S(t)x2 in proposition 5.14, and deduce that kS(t)x1 − S(t)x2 kX ≤ c2 kS(t)x1 − wkX .

(5.30)

Recalling now (5.23), which is a consequence of the exponential tracking property (5.7), and that π1 S(t)x1 = π1 w, we deduce from (5.30) that kS(t)x1 − S(t)x2 kX ≤ c2 kS(t)x1 − wkX = c2 kπ2 (S(t)x1 − w)kX = c2 kπ2 S(t)x1 − m(π1 S(t)x1 )kX ≤ c2Cd(x1 , M)e−ηt ≤ c2Ckx1 − x2 kX e−ηt ,

(5.31)

where the last step follows because x2 ∈ M. In conclusion, we deduce that if S admitted an inertial manifold M of the form (5.16), with m ∈ GL0 , L0 < L, then condition (5.29) would imply that there is K > 0, independent of x1 , x2 and t, such that kS(t)x1 − S(t)x2 kX ≤ Kkx1 − x2 kX e−ηt .

(5.32)

5.3

Geometric Assumptions on the Semiflow

193

At this point, we realize that conditions (5.29) and (5.32) can be formulated independently of the requirement that x1 , x2 ∈ M. This motivates the following DEFINITION 5.15 Let η, L > 0, and CL as in (5.24). The semiflow S satisfies the DECAY PROPERTY with parameters η and L if there is K > 0 with the property that whenever x1 , x2 ∈ X and t > 0 are such that (5.29) holds, then (5.32) also holds. Thus, the decay property means that the difference of motions outside the cone CL decays exponentially. For future reference, we remark that, as an immediate consequence of proposition 5.13, if (5.29) and (5.32) hold for a specific t > 0, then they both hold for all t 0 ∈ [0,t]. Finally, we combine the cone invariance and the decay properties into DEFINITION 5.16 In the same conditions of definition 5.15, the semiflow S satisfies the STRONG SQUEEZING PROPERTY with parameters η and L if it satisfies the cone invariance property with parameter L, and the decay property with parameters η and L. The strong squeezing property was first introduced for the Kuramoto-Sivashinski equations in Foias, Nicolaenko, Sell and Temam, [FNST85, FNST88]. An abstract version of this property was developed in Foias, Sell and Titi, [FST89]; other formulations can be found e.g. in Temam, [Tem88], Constantin, Foias, Nicolaenko and Temam, [CFNT89], Robinson, [Rob93], and Jones and Titi, [JT96]. The term “strong” refers to the fact that, in contrast e.g. to the discrete squeezing property, we require that a cone invariance property hold as well.

5.3.3 Consequences of the Decay Property In this section we briefly report two immediate consequences of the decay property. We first show that if a semiflow S admits an attractor, and satisfies the decay property, the attractor is smooth, in the sense that it is imbedded into a Lipschitz manifold. PROPOSITION 5.17 Assume the semiflow S satisfies the decay property with parameters L and η, and admits a global attractor A. There exists then a function m ∈ GL such that A ⊆ graph(m). PROOF Let x1 and x2 ∈ A. Since A is invariant, for each t > 0 there are y1 and y2 ∈ A such that xi = S(t)yi , i = 1, 2. If x1 − x2 ∈ / CL , the decay property (5.32) implies that kx1 − x2 kX = kS(t)y1 − S(t)y2 kX ≤ Kky1 − y2 kX e−ηt .

194

5

Inertial Manifolds

Since A is bounded, we deduce that kx1 − x2 kX ≤ R e−ηt , for some R independent of t. Letting t → +∞, we conclude that x1 = x2 . Thus, we can define a function m : π1 A → X2 by m(π1 x) := π2 x ,

x ∈ A.

Indeed, let y ∈ A be such that π1 y = π1 x. Then: if y − x ∈ / CL , we have shown that y = x; if y − x ∈ CL , then π2 y = π2 x so that, again, x = y. The function m so defined is Lipschitz continuous, as a consequence of proposition 5.10. Indeed, if x, y ∈ graph(m), and x 6= y, then x − y ∈ CL by the first part of this proof. We can then conclude by extending m to a bounded function, defined on all of X1 , and having the same Lipschitz constant L. Next, we report a technical result which will be used in the next section. PROPOSITION 5.18 Assume S is a continuous semiflow on X , satisfying the strong squeezing property with parameters L and η. Assume that t1 and t2 are such that 0 < t1 < t2 and there are x1 , x2 ∈ X1 such that π1 S(t1 )x1 = π1 S(t2 )x2 .

(5.33)

There is then C1 > 0, independent of t1 , t2 , x1 , x2 , such that kπ2 (S(t1 )x1 − S(t2 )x2 )kX ≤ C1 kπ2 (x1 − S(t2 − t1 )x2 )kX e−ηt1 .

(5.34)

PROOF If S(t1 )x1 − S(t2 )x2 ∈ CL , (5.33) implies that π2 (S(t1 )x1 − S(t2 )x2 ) = 0; thus, (5.34) holds trivially. Otherwise, we can write that S(t1 )x1 − S(t2 )x2 = S(t1 )x1 − S(t1 )S(t2 − t1 )x2 ∈ / CL , and proposition 5.13 implies that also x1 − S(t2 − t1 )x2 ∈ / CL . By the decay property (5.32), with t = t1 and x2 replaced by S(t2 − t1 )x2 , we compute then kπ2 (S(t1 )x1 − S(t2 )x2 ) kX = kS(t1 )x1 − S(t2 )x2 kX = kS(t1 )x1 − S(t1 )S(t2 − t1 )x2 kX ≤ Kkx1 − S(t2 − t1 )x2 kX e−ηt1 ≤ K kπ1 (x1 − S(t2 − t1 )x2 ) kX + kπ2 (x1 − S(t2 − t1 )x2 ) kX e−ηt1 ≤ K L1 + 1 kπ2 (x1 − S(t2 − t1 )x2 ) kX e−ηt1 , the last step as in (5.28), because x1 − S(t2 − t1 )x2 ∈ / CL . Thus, (5.34) follows.

5.4

Strong Squeezing Property and Inertial Manifolds

195

5.4 Strong Squeezing Property and Inertial Manifolds We are now in a position to show that if the semiflow S satisfies the strong squeezing property, it admits an inertial manifold of the form (5.16), which we can construct by Hadamard’s graph transformation method. In the sequel, we closely follow Robinson, [Rob01].

5.4.1 Surjectivity and Uniform Boundedness For our first result, we need DEFINITION 5.19 Let S be a continuous semiflow on X ; let GL be as in (5.19), and Φ ⊆ GL . Let π1 and π2 be as in (5.32). Then: 1. S satisfies the SURJECTIVITY PROPERTY with respect to Φ if for all ϕ ∈ Φ and t ≥ 0, π1 S(t) graph(ϕ) = X1 .

(5.35)

2. S satisfies the UNIFORM BOUNDEDNESS PROPERTY with respect to Φ if the sets π2 S(t) graph(ϕ) are bounded in X , uniformly with respect to t ≥ 0 and ϕ ∈ Φ. We have then: PROPOSITION 5.20 Let S be a semiflow on X . Assume S satisfies the cone invariance property with parameter L, and the surjectivity and uniform boundedness properties with respect to some subset Φ ⊆ GL . Then for each ϕ0 ∈ Φ and t > 0 there exists a uniquely determined function ϕt ∈ GL such that graph(ϕt ) = S(t) graph(ϕ0 ) .

(5.36)

PROOF 1. Fix ϕ0 ∈ Φ and t > 0. Since S satisfies the surjectivity property, (5.35) implies that π1 S(t) graph(ϕ0 ) = X1 . Thus, for each ξ ∈ X1 there is x ∈ S(t) graph(ϕ0 ) such that π1 x = ξ .

(5.37)

We claim that x is uniquely determined by ξ (and, of course, by t and ϕ0 ). Indeed, suppose x0 ∈ S(t) graph(ϕ0 ) is also such that π1 x0 = ξ as in (5.37). Let z and z0 ∈

196

5

Inertial Manifolds

graph(ϕ0 ) be such that x = S(t)z and x0 = S(t)z0 . Since ϕ0 ∈ GL , proposition 5.10 implies that z − z0 ∈ CL . Thus, the cone invariance property (5.26) implies that also x − x0 ∈ CL , i.e. kπ2 (S(t)z − S(t)z0 )kX ≤ Lkπ1 (S(t)z − S(t)z0 )kX .

(5.38)

But since π1 S(t)z = π1 x = ξ = π1 x0 = π1 S(t)z0 , (5.38) shows that π2 S(t)z = π2 S(t)z0 . Hence, S(t)z = S(t)z0 , i.e. x = x0 as claimed. 2. From part (1) it follows that we can unambiguously define a function κ from R≥0 × X1 × Φ into S(t) graph(ϕ0 ), by (t, ξ , ϕ0 ) 7→ x := κ(t, ξ , ϕ0 ) .

(5.39)

We can then define a map ϕt : X1 → X2 by ϕt (ξ ) := π2 κ(t, ξ , ϕ0 ) ,

ξ ∈ X1 .

That is, ϕt (ξ ) = π2 x, where x ∈ X is the unique element in S(t) graph(ϕ0 ) such that π1 x = ξ . In other words, π1 κ(t, ξ , ϕ0 ) = ξ , π2 κ(t, ξ , ϕ0 ) = ϕt (ξ ) .

(5.40)

Note that, by construction, x = ξ + ϕt (ξ ) ∈ graph(ϕt ) . We proceed then to show (5.36). Indeed, let first z ∈ graph(ϕt ). Then, z = η + ϕt (η), for some η ∈ X1 . By (5.40), z = κ(t, η, ϕ0 ); thus, by (5.39), z ∈ S(t) graph(ϕ0 ). Conversely, let z ∈ S(t) graph(ϕ0 ), and set ξ := π1 z. Then, ξ uniquely determines x := κ(t, ξ , ϕ0 ) ∈ S(t) graph(ϕ0 ) = graph(ϕt ), with π1 x = ξ . Since z ∈ S(t) graph(ϕ0 ), and ξ = π1 z, the stated uniqueness implies that x = z. Thus, z = x ∈ graph(ϕt ), and (5.36) holds. 3. Finally, we show that ϕt ∈ GL , i.e. that ϕt : X1 → X2 is bounded and Lipschitz continuous, with the same Lipschitz constant L of ϕ0 . To show that ϕt is bounded, we must exhibit R1 > 0 such that kϕt kG = sup kϕt (ξ )kX = sup kπ2 κ(t, ξ , ϕ0 )kX ≤ R1 . ξ ∈X1

(5.41)

ξ ∈X1

Since S satisfies the uniform boundedness property with respect to Φ, there is R > 0 such that for all t > 0, ϕ ∈ Φ, and y ∈ S(t) graph(ϕ), kπ2 ykX ≤ R .

(5.42)

But since ϕ0 ∈ Φ and κ(t, ξ , ϕ0 ) ∈ S(t) graph(ϕ0 ), there is y ∈ graph(ϕ0 ) such that κ(t, ξ , ϕ0 ) = S(t)y. Hence, (5.41) follows from (5.42), with R1 = R. To show that ϕt is Lipschitz continuous, given ξ1 and ξ2 ∈ X1 , let xi = ξi + ϕt (ξi ), i = 1, 2. Then xi ∈

5.4

197

Strong Squeezing Property and Inertial Manifolds

graph(ϕt ) = S(t) graph(ϕ0 ); thus, xi = S(t)yi for some yi = ηi + ϕ0 (ηi ) ∈ graph(ϕ0 ). Since ϕ0 ∈ GL , y1 −y2 ∈ CL ; hence, by the cone invariance property (5.26), x1 −x2 ∈ CL as well. Thus, kϕt (ξ1 ) − ϕt (ξ2 )kX = kπ2 (x1 − x2 )kX ≤ Lkπ1 (x1 − x2 )kX = Lkξ1 − ξ2 kX . This shows that ϕt ∈ GL , and concludes the proof of proposition 5.20. As a consequence of proposition 5.20, if for t > 0 we set Mt := graph(ϕt ) ,

(5.43)

Mt = S(t)M0 .

(5.44)

then (5.36) implies that

5.4.2 Construction of the Inertial Manifold We can finally proceed to implement Hadamard’s construction of an inertial manifold. THEOREM 5.21 Let S be a continuous semiflow on X . Assume that S satisfies the strong squeezing property with parameters L and η, as well as the surjectivity and uniform boundedness properties with respect to the set Φ := {0} ⊂ GL . Then S admits an inertial manifold M ⊆ X , of the form (5.16). More precisely, there is m ∈ GL such that M := graph(m) is an inertial manifold for S. Moreover, for all x ∈ X and t ≥ 0, d(S(t)x, M) ≤ C1 (d(x, M) + 2kmkGL ) e−ηt ,

(5.45)

where C1 is as in (5.34). PROOF We shall use the graph transformation method, and proceed in three steps. 1. Let ϕ0 ≡ 0. Then, obviously, M0 := graph(ϕ0 ) = X1 ,

(5.46)

which is a trivial flat manifold. In accord to (5.44), we follow the evolution of this flat manifold. Proposition 5.20 shows that each Mt is again the graph of a function ϕt ∈ GL . Using the decay property and the uniform boundedness property, we will show that, as t → +∞, the functions (ϕt )t ≥0 converge to a function m ∈ GL , whose graph M is the desired inertial manifold for S. 2. Recalling (5.43) and (5.46), by proposition 5.20 we have that for all t ≥ 0 there is ϕt ∈ GL such that Mt = graph(ϕt ) = S(t) graph(ϕ0 ) = S(t)X1 .

(5.47)

198

5

Inertial Manifolds

Since S satisfies the uniform boundedness property with respect to Φ = {0}, and graph(ϕ0 ) = X1 , there is R > 0 such that (5.42) holds for all t > 0 and y ∈ S(t)X1 . Given ξ ∈ X1 and t > 0, let x = κ(t, ξ , 0). Then, x ∈ S(t) graph(ϕ0 ) = S(t)X1 , and π2 x = ϕt (ξ ). Thus, by (5.42), kϕt kG = sup kϕt (ξ )kX = sup kπ2 κ(t, ξ , 0)kX ≤ R . ξ ∈X1

(5.48)

ξ ∈X1

Fix t1 and t2 , with 0 < t1 < t2 , and ξ ∈ X1 . For i = 1, 2, let xi := ξ + ϕti (ξ ) ∈ Mti . By (5.47), there are ξ1 , ξ2 ∈ X1 such that xi = S(ti )ξi . Since π1 S(t1 )ξ1 = π1 S(t2 )ξ2 = ξ , (5.33) holds, and proposition 5.18 implies that, by (5.34), kπ2 (S(t1 )ξ1 − S(t2 )ξ2 )kX ≤ C1 kπ2 (ξ1 − S(t2 − t1 )ξ2 )kX e−ηt1 .

(5.49)

Now, since ξ1 ∈ X1 , π2 ξ1 = 0; since also S(t2 − t1 )ξ2 ∈ S(t2 − t1 )X1 = Mt2 −t1 , it follows that π2 S(t2 − t1 )ξ2 = ϕt2 −t1 (π1 S(t2 − t1 )ξ2 ) . Moreover, π2 (S(t1 )ξ1 − S(t2 )ξ2 ) = π2 (x1 − x2 ) = ϕt1 (ξ ) − ϕt2 (ξ ) . Hence, we deduce from (5.49) and (5.48) that kϕt1 (ξ ) − ϕt2 (ξ )kX ≤ C1 kϕt2 −t1 (π1 S(t2 − t1 )ξ2 )kX e−ηt1 ≤ C1 Re−ηt1 .

(5.50)

Since ξ ∈ X1 and t1 , t2 are arbitrary, this implies that (ϕt )t ≥0 is a Cauchy set in GL : in fact, given ε > 0, by (5.50) we have that kϕt − ϕθ kGL = sup kϕt (ξ ) − ϕθ (ξ )k < ε ξ ∈X1

for all t and θ such that n o min{t, θ } ≥ T := max 0, ln Cε1 R . Since GL is a complete metric space, we conclude that, as t → +∞, (ϕt )t ≥0 converges to a function m ∈ GL . As a consequence of (5.48), m is also bounded, with kmkGL ≤ R. 3. We proceed then to show that the set M := graph(m) is the desired inertial manifold of S. Thus, we need to show that M is positively invariant, and exponentially attracting. 3.1. The positive invariance of M is a consequence of the uniform convergence ϕt → m. Indeed, let x = ξ + m(ξ ) ∈ M. For t ≥ 0, let xt := ξ + ϕt (ξ ) ∈ Mt . Then, as t → +∞, xt → ξ + m(ξ ) = x. Fix τ > 0: then, by (5.44), S(τ)xt ∈ S(τ)Mt = S(τ)S(t)M0 = S(τ + t)M0 = Mτ+t ;

5.4

199

Strong Squeezing Property and Inertial Manifolds

therefore, S(τ)xt = π1 S(τ)xt + ϕτ+t (π1 S(τ)xt ) .

(5.51)

Since S(τ) is continuous on X , S(τ)xt → S(τ)x as t → +∞; on the other hand, the right side of (5.51) converges to π1 S(τ)x + m(π1 S(τ)x). It follows that S(τ)x = π1 S(τ)x + m(π1 S(τ)x) ∈ graph(m) = M . Since x ∈ M and τ > 0 are arbitrary, this means that M is positively invariant. 3.2. We now show that M is exponentially attracting, and (5.45) holds. Fix x ∈ X \ M, and t > 0. Let ξ := π1 S(t)x and, for θ > t, y := ξ + ϕθ (ξ ). Then y ∈ Mθ = S(θ )X1 , so there is z ∈ X1 such that y = S(θ )z. Since π1 S(t)x = π1 S(θ )z = ξ , by (5.34) of proposition 5.18 we have kπ2 (S(t)x − S(θ )z)kX ≤ C1 kπ2 (x − S(θ − t)z)kX e−ηt . Let w := S(θ − t)z. Then w ∈ Mθ −t , so w = α + ϕθ −t (α) for some α ∈ X1 . Since π2 S(θ )z = π2 y = ϕθ (ξ ), we can estimate d(S(t)x, M) ≤ kS(t)x − (ξ + m(ξ ))kX = kπ2 S(t)x − m(ξ )kX = lim kπ2 S(t)x − ϕθ (ξ )kX = lim kπ2 (S(t)x − S(θ )z)kX θ →+∞

θ →+∞

≤ C1 lim kπ2 (x − S(θ − t)z)kX e−ηt θ →+∞

= C1 lim kπ2 x − ϕθ −t (α)kX e−ηt θ →+∞

= C1 kπ2 x − m(α)kX e−ηt . Given then arbitrary u = β + m(β ) ∈ M, we deduce that d(S(t)x, M) ≤ C1 (kπ2 x − m(β )kX + km(β ) − m(α)kX )e−ηt = C1 (kπ2 (x − u)kX + km(β ) − m(α)kX )e−ηt ≤ C1 (kx − ukX + 2kmkGL ) e−ηt . Taking inf with respect to u ∈ M, we conclude that d(S(t)x, M) ≤ C1 (d(x, M) + 2kmkGL )e−ηt , which is (5.45). This concludes the proof of theorem 5.21. We remark that the inertial manifold M so constructed is not compact. This is because M is the graph of a function m over X1 , which is not bounded. To obtain a compact inertial manifold, we can apply proposition 5.9, and intersect M with a compact, positively invariant absorbing set. On the other hand, since m is Lipschitz continuous, proposition 2.61 implies that M has finite fractal dimension, with dimF (M) = dim X1 .

200

5

Inertial Manifolds

We conclude this section with a property that characterizes the convergence of the sets S(t)X1 to the inertial manifold M = graph(m) constructed in the previous section. PROPOSITION 5.22 Let (ϕt )t ≥0 be the family of functions in GL constructed in proposition 5.20 from the initial function ϕ0 ≡ 0, as in theorem 5.21. Let m := lim ϕt ∈ GL . t →+∞

Then the semiflow S satisfies the surjectivity property with respect to the set Φ1 := (ϕt )t ≥0 ∪ {m} , and lim dist(graph(ϕt ), graph(m)) = 0 ,

t →+∞

(5.52)

where dist is the distance defined in (2.3). PROOF 1. Recalling (5.35), we need to prove that for all t, θ ≥ 0, π1 S(t) graph(ϕθ ) = X1 , π1 S(t) graph(m) = X1 .

(5.53)

Obviously, π1 S(t) graph(ϕθ ) ⊆ X1 and π1 S(t) graph(m) ⊆ X1 . To prove the inverse inclusions, fix ξ ∈ X1 and t, θ > 0. Let first y := ξ + ϕt+θ (ξ ). Then, by (5.36) and (5.46), y ∈ graph(ϕt+θ ) = S(t + θ )X1 = S(t)S(θ )X1 ; thus, there is xθ ∈ S(θ )X1 such that y = S(t)xθ . But then xθ ∈ graph(ϕθ ), and π1 S(t)xθ = ξ . This proves the first of (5.53). To prove the second, we proceed as before. Then, since xθ ∈ graph(ϕθ ), there is α ∈ X1 such that xθ = α + ϕθ (α). Let z := α + m(α). Since the convergence ϕθ → m as θ → +∞ is uniform, xθ → z as θ → +∞. Thus, since both π1 and S(t) are continuous, kξ − π1 S(t)zkX = kπ1 (S(t)xθ − S(t)z)kX → 0 as well. It follows that π1 S(t)z = ξ ; since z ∈ M, this proves the second of (5.53). 2. To prove (5.52), recalling (2.3), (5.36) and (5.46) we need to show that, as t → +∞, ∂ (S(t)X1 , M) → 0 , ∂ (M, S(t)X1 ) → 0 .

(5.54)

Given at = α + ϕt (α) ∈ graph(ϕt ) = S(t)X1 , let b := α + m(α) ∈ M. Then 0 ≤ d(at , M) ≤ kat − bkX = kϕt (α) − m(α)kX → 0 .

(5.55)

5.5

A Modification

201

Moreover, since the convergence ϕt → m is uniform with respect to α ∈ X1 , (5.55) implies that, in fact, sup

d(at , M) = ∂ (S(t)X1 , M) → 0 .

at ∈S(t)X1

This is the first of (5.54); the second is proven analogously. REMARK 5.23 The second of (5.53) implies that for each ξ ∈ X1 and t ≥ 0 we can determine x ∈ M such that π1 S(t)x = ξ .

(5.56)

The following result shows that this element x can be determined uniquely, if S satisfies the backward uniqueness property (recall definition 3.14). An instance when this property is satisfied is given in proposition 5.27 below. PROPOSITION 5.24 In the same conditions of proposition 5.22, assume that S satisfies the backward uniqueness property. Then the element x in (5.56) is uniquely determined by ξ (and t). PROOF Assume that y ∈ M is also such that π1 S(t)y = ξ . Then, since m ∈ GL , x − y ∈ CL . By the cone invariance property, also S(t)x − S(t)y ∈ CL . But since π1 S(t)x = ξ = π1 S(t)y, it follows that S(t)x = S(t)y. The backward uniqueness property implies then that x = y.

5.5 A Modification In this section we present a second result, whereby the existence of an inertial manifold of the form (5.16) follows from a modified version of the strong squeezing property, related to a modification of the decay property. The main interest of this result lies in an improved estimate of the exponential convergence of the orbits to the manifold, which involves only the initial distance d(x, M) of the orbit to the manifold.

5.5.1 The Modified Strong Squeezing Property In analyzing the proof of theorem 5.21, we realize that the only point where the decay property was used in an essential way was in the proof of proposition 5.18. We further observe that its assumption (5.33) can be rewritten as π1 S(t1 )u = π1 S(t1 )v ,

(5.57)

202

5

Inertial Manifolds

for u = x1 and v = S(t2 − t1 )x2 . This is our starting point for the introduction of the modified strong squeezing property. Again, we assume for the moment that S does admit an inertial manifold M of the form (5.16), and proceed to deduce other natural geometric properties of the semiflow. We go back to the last step of (5.31), with the difference that we now use the estimate d(x1 , M) ≤ kx1 − zkX , with z := π1 x1 + m(π1 x1 ) ∈ M. Together with (5.31), this yields kS(t)x1 − S(t)x2 kX ≤ CC1 kx1 − zkX e−ηt .

(5.58)

At this point, the right side of (5.58) still depends on the function m, which of course is the very function we want to determine, via the choice of z. To remove this dependency, we single out two properties of z that can be expressed without the explicit knowledge of m; namely, that π1 z = π1 x1 , and that x2 − z ∈ CL . Of these, the first follows from the definition of z, and the second would follow from proposition 5.10, recalling that both z and x2 ∈ M (the latter as assumed before (5.29)). Thus, we are led to introduce the set Y(x1 , x2 ) := {u ∈ X : π1 u = π1 x1 , u − x2 ∈ CL } ,

(5.59)

to which z would belong. We further note that if we also require that kS(t)x1 − S(t)x2 kX ≤ CC1

inf

u∈Y (x1 ,x2 )

kx1 − ukX e−ηt ,

(5.60)

then (5.58) would follow from (5.60), since z ∈ Y(x1 , x2 ). Finally, recalling (5.57), we further modify (5.60) by assuming that π1 S(t)x1 = π1 S(t)x2 , so that the left side of (5.60) reduces to kπ2 (S(t)x1 − S(t)x2 )kX . In conclusion, renaming t = T , we have seen that if S admitted an inertial manifold M of the form (5.16), with m ∈ GL , then the conditions π1 S(T )x1 = π1 S(T )x2 ,

x1 , x2 ∈ X , T > 0 ,

(5.61)

would imply that there is K > 0 such that kπ2 (S(T )x1 − S(T )x2 )kX ≤ K

inf

u∈Y (x1 ,x2 )

kx1 − ukX e−ηT .

(5.62)

Our last step is to require that (5.62) holds not only for t = T , but also for all t ∈ [0, T ]. That is, we give DEFINITION 5.25 Let L, η > 0, and CL be as in (5.24). The semiflow S satisfies the MODIFIED DECAY PROPERTY with parameters L and η if there is K > 0 such that whenever x1 , x2 ∈ X and T > 0 are such that (5.61) holds, and π2 S(T )x1 6= π2 S(T )x2 ,

(5.63)

5.5

203

A Modification

then for all t ∈ [0, T ], kπ2 (S(t)x1 − S(t)x2 )kX ≤ K

inf

u∈Y (x1 ,x2 )

kx1 − ukX e−ηt ,

(5.64)

where the set Y(x1 , x2 ) is defined as in (5.59). Combining the cone invariance property with the modified decay property we finally arrive at DEFINITION 5.26 In the same conditions of definition 5.25, the semiflow S satisfies the MODIFIED STRONG SQUEEZING PROPERTY with parameters L and η if it satisfies the cone invariance property with parameter L and the modified decay property with parameters L and η.

5.5.2 Consequences of the Modified Strong Squeezing Property In this section we present some immediate consequences of the modified strong squeezing property. We assume that L, η and CL are as in the previous definitions, and that S satisfies the modified strong squeezing property with parameters L and η. PROPOSITION 5.27 (Backward Uniqueness) Assume that x1 , x2 ∈ X and T > 0 are such that (5.61) and (5.63) hold. Then for all t ∈ [0, T ], kS(t)x1 − S(t)x2 kX ≤ K 1 + L1

inf

u∈Y (x1 ,x2 )

kx1 − ukX e−ηt .

(5.65)

In particular, if x1 − x2 ∈ CL , then x1 = x2 . PROOF If (5.61) and (5.63) hold, it follows that for all t ∈ [0, T ], S(t)x1 − S(t)x2 ∈ / CL . Indeed, if there were θ ∈ [0, T ] such that S(θ )x1 −S(θ )x2 ∈ CL , the cone invariance property would imply that S(T )x1 − S(T )x2 ∈ CL as well. But then (5.61) would imply that π2 S(T )x1 = π2 S(T )x2 , contradicting (5.63). Now, since the condition S(t)x1 − S(t)x2 ∈ / CL is equivalent to the inequality kπ2 (S(t)x1 − S(t)x2 )kX > Lkπ1 (S(t)x1 − S(t)x2 )kX ,

(5.66)

(5.65) follows from (5.66) and (5.64). In particular, if x1 −x2 ∈ CL , then x1 ∈ Y(x1 , x2 ). Hence, if x1 − x2 ∈ CL , the right-hand side of (5.65), and therefore also the difference S(t)x1 − S(t)x2 , vanishes for all t ∈ [0, T ] . In particular, from proposition 5.24 we deduce that if S satisfies the strong squeezing property, the element x ∈ M defined by (5.56) is uniquely determined by ξ ∈ X1 .

204

5

Inertial Manifolds

PROPOSITION 5.28 (Contractivity) Assume that ϕ1 , ϕ2 ∈ GL , and x1 ∈ graph(ϕ1 ), x2 ∈ graph(ϕ2 ), T > 0 are such that (5.61) and (5.63) hold. There exists C > 0, independent of x1 , x2 , such that for all t ∈ [0, T ], kπ2 (S(t)x1 − S(t)x2 )kX ≤ Ckϕ1 (π1 x1 ) − ϕ2 (π1 x2 )kX e−ηt .

(5.67)

PROOF Let y := π1 x1 + ϕ2 (π1 x2 ). Then, since x2 ∈ graph(ϕ2 ), π2 (y − x2 ) = ϕ2 (π1 x2 ) − ϕ2 (π1 x2 ) = 0 . Hence, y − x2 ∈ CL (trivially), and since π1 y = π1 x1 , it follows that y ∈ Y(x1 , x2 ). Since also x1 − y = π2 x1 − π2 y = ϕ1 (π1 x1 ) − ϕ2 (π1 x2 ) , (5.67) follows from (5.64). PROPOSITION 5.29 (Asymptotic Phases) Let m ∈ GL , and assume that x1 ∈ X , x2 ∈ graph(m) and T > 0 are such that (5.61) and (5.63) hold. Then for all t ∈ [0, T ], kπ2 (S(t)x1 − S(t)x2 )kX ≤ Ckπ2 x1 − m(π1 x1 )kX e−ηt .

(5.68)

PROOF Let y := π1 x1 + m(π1 x1 ). Then y ∈ graph(m), and y − x2 ∈ CL by proposition 5.10. Hence, y ∈ Y(x1 , x2 ). Since also x1 − y = π2 x1 − m(π1 x1 ), (5.68) follows from (5.64). We can then conclude, recalling definition 5.7.

5.5.3 Construction of the Inertial Manifold, 2 We can now show that if the semiflow S satisfies the modified strong squeezing property, it admits an inertial manifold of the form (5.16), which satisfies an exponential estimate involving only the initial distance of the orbits to the manifold. We claim: THEOREM 5.30 Let S be a continuous semiflow on X , which satisfies the modified strong squeezing property with parameters L and η, as well as the surjectivity and uniform boundedness properties with respect to the set Φ = {0}. Then S admits an inertial manifold M ⊆ X of the form (5.16). More precisely, there is m ∈ GL , such that the set M := graph(m) is an inertial manifold for S, having the exponential tracking property (5.5). That is, for all x ∈ X \ M, there is z ∈ M such that for all t ≥ 0, kS(t)x − S(t)zkX ≤ C2 kπ2 x − m(π1 x)kX e−ηt ,

(5.69)

with C2 independent of x, z and t. Consequently, M is exponentially attracting.

5.5

A Modification

205

PROOF We start as in the proof of theorem 5.21. Keeping the same notations, we see that we can proceed exactly in the same way up to the proof that M is positively invariant, except that now we cannot use proposition 5.18 to obtain (5.49). However, we can instead use proposition 5.28, with x1 = ξ1 ∈ X1 = graph(ϕ0 ), and x2 = S(t2 −t1 )ξ2 ∈ S(t2 −t1 )X1 = graph(ϕt2 −t1 ); note that π1 S(t1 )x1 = π1 S(t1 )x2 = ξ . Thus, by (5.67) we obtain kπ2 (x1 − x2 )kX ≤ kπ2 (S(t1 )ξ1 − S(t2 )ξ2 )kX = kπ2 (S(t1 )ξ1 − S(t1 )S(t2 − t1 )ξ2 )kX ≤ C1 kϕ0 (π1 ξ1 ) − ϕt2 −t1 (π1 S(t2 − t1 )ξ1 )kX e−ηt1 . Since ϕ0 (π1 ξ1 ) = 0, (5.49) follows. To show that M has the exponential tracking property, given x ∈ X \M we consider the set Jx := {t ≥ 0 : (∃ z ∈ M : π1 S(t)z = π1 S(t)x, π2 S(t)z 6= π2 S(t)x)} . This set is not empty, since t = 0 ∈ Jx . To see this, let z := π1 x + m(π1 x). Then π1 z = π1 x, but π2 z = m(π1 x) 6= π2 x, since z ∈ M but x ∈ / M. Consider first the case when sup Jx =: t < +∞. Then for all t ≥ t there is z ∈ M such that S(t)z = S(t)x. Then, (5.69) is trivial if t ≥ t, while if t ≤ t it is a consequence of the modified decay property. Indeed, in this case the left side of (5.69) reduces to kπ2 (S(t)x − S(t)z)kX , which we can estimate by (5.64). In particular, we can take u = π1 x + m(π1 x) which is in Y(x, z), because both u and z ∈ M. Hence, from (5.64), kπ2 (S(t)x − S(t)z)kX ≤ Kkx − ukX e−ηt , and (5.69) follows, because of the choice of u. If instead sup Jx = +∞, we can find a sequence (tk )k∈N such that tk → +∞ monotonically, and a corresponding sequence (zk )k∈N ⊂ M such that for all k ∈ N, π1 S(tk )zk = π1 S(tk )x and π2 S(tk )zk 6= π2 S(tk )x . Since m ∈ GL , we can apply proposition 5.29 and (5.66), to deduce that for each k ∈ N and all t ∈ [0,tk ], L kπ1 (S(t)x − S(t)zk )kX ≤ kπ2 (S(t)x − S(t)zk )kX ≤ Ckπ2 x − m(π1 x)kX e−ηt . (5.70) In particular, for t = 0 this implies that for all k ∈ N kπ1 x − π1 zk kX ≤ CL kπ2 x − m(π1 x1 )kX =: ρ . This means that each ζk := π1 zk is in the closed ball of X1 of center π1 x and radius ρ. Since X1 has finite dimension, this ball is compact. Thus, there is a subsequence of (tk )k∈N , still denoted (tk )k∈N , such that the corresponding subsequence (ζk )k∈N

206

5

Inertial Manifolds

converges to a limit ζ ∈ X1 , with kπ1 x − ζ kX ≤ ρ. Let z := ζ + m(ζ ). Then for each k ∈ N and all t ∈ [0,tk ], recalling (5.70) we can estimate kS(t)x − S(t)zkX ≤ kS(t)x − S(t)zk kX + kS(t)zk − S(t)zkX (5.71) ≤ 1 + L1 kπ2 (S(t)x − S(t)zk )kX + kS(t)zk − S(t)zkX −ηt 1 ≤ C 1 + L kπ2 x − m(π1 x1 )kX e + kS(t)zk − S(t)zkX . Since m is continuous, and z, zk ∈ M for all k ∈ N, we have that zk → z. Since S(t) is continuous, it follows that also S(t)zk → S(t)z. Thus, there is k0 ∈ N (depending on t), such that if k ≥ k0 , kS(t)zk − S(t)zkX ≤ C 1 + L1 kπ2 x − m(π1 x1 )kX e−ηt .

(5.72)

Consequently, (5.69) follows from (5.71) and (5.72) (recall that z ∈ M). From (5.6) we know that the exponential tracking property implies that M is exponentially attracting; thus, the proof of theorem 5.30 is complete. We remark that the inertial manifold so constructed enjoys all the properties described in section 5.4.2.

5.5.4 Comparison of the Squeezing Properties We conclude this section with a comparison of the strong squeezing property and the discrete squeezing property introduced in definition 4.3, in the context of exponential attractors. We also compare the strong squeezing property with the modified strong squeezing property. At first, we note that if the semiflow S satisfies the strong squeezing property of definition 5.16 with parameters η and L, and L ≤ 1, then S also satisfies the discrete squeezing property of definition 4.3, with B = X , PN = π1 , XN = X1 . Indeed, comparing (4.9) with (5.24) we have CP = CPN = C1 ⊆ CL for all L ∈ ]0, 1]. Given γ ∈ ]0, 12 [, choose t∗ = T so large that Ke−ηT ≤ γ. Then, if u, v ∈ X are such that S(T )u − S(T )v ∈ / C1 , and therefore also S(T )u − S(T )v ∈ / CL for L ≤ 1, (5.32) implies that kS(T )u − S(T )vkX ≤ Ke−ηT ku − vkX ≤ γku − vkX . This means that the operator S∗ = S(T ) satisfies (4.11); hence, the discrete squeezing property holds. Next, to illustrate the relationship between the strong squeezing property and the modified strong squeezing property, we return to the observation we made at the beginning of section 5.3.2: namely, that we can often assume that the semiflow S also satisfies a second cone invariance property. In this case, however, we assume that this second cone invariance property involves a cone CL0 , with parameter L0 > L (instead of L0 < L, as in proposition 5.14). In these conditions, we claim:

5.5

207

A Modification

PROPOSITION 5.31 Assume the continuous semiflow S satisfies the strong squeezing property with parameters L and η. Suppose further that there exists L0 > L such that S also satisfies a second cone invariance property with parameter L0 . Then S satisfies the modified strong squeezing property, with parameters L and η. PROOF Let x1 , x2 ∈ X , and T > 0 be such that π1 S(T )x1 = π1 S(T )x2 , π2 S(T )x1 6= π2 S(T )x2 , as required in (5.61) and (5.63). Then, we have that both S(T )x1 − S(T )x2 6∈ CL , S(T )x1 − S(T )x2 6∈ CL0 , for, otherwise, definition (5.24) would imply that π2 S(T )x1 = π2 S(T )x2 . Proposition 5.13 implies then that S(t)x1 − S(t)x2 6∈ CL , S(t)x1 − S(t)x2 6∈ CL0 for all t ∈ [0, T ]. Consequently, for t ∈ [0, T ], Lkπ1 (S(t)x1 − S(t)x2 )kX < L0 kπ1 (S(t)x1 − S(t)x2 )kX < kπ2 (S(t)x1 − S(t)x2 )kX .

(5.73)

Let x3 ∈ Y(x1 , x2 ) (the set introduced in (5.59)). Then π1 x3 = π1 x1 and x3 − x2 ∈ CL , and (5.24) implies that kπ2 (x2 − x3 )kX ≤ Lkπ1 (x2 − x1 )kX . This, together with (5.73) for t = 0, implies L0 kπ1 (x1 − x2 )kX ≤ kπ2 (x1 − x2 )kX ≤ kπ2 (x1 − x3 )kX + kπ2 (x2 − x3 )kX ≤ kπ2 (x1 − x3 )kX + Lkπ1 (x1 − x2 )kX . (5.74) Hence, kπ1 (x1 − x2 )kX ≤

1 L0 −L kπ2 (x1 − x3 )kX

,

(5.75)

L0 L0 −L kπ2 (x1 − x3 )kX

.

(5.76)

and, replacing back into (5.74), kπ2 (x1 − x2 )kX ≤

By the decay property (5.32), recalling (5.75) and (5.76), kS(t)x1 − S(t)x2 kX ≤ Kkx1 − x2 kX e−ηt ≤ K( L10 + 1)kπ2 (x1 − x2 )kX e−ηt ≤

K(1+L0 ) −ηt L0 −L kπ2 (x1 − x3 )kX e

.

(5.77)

208

5

Inertial Manifolds

Since π1 x3 = π1 x1 , we also have kπ2 x1 − π1 x3 kX = kπ2 x1 + π1 x1 − π1 x3 − π2 x3 kX = kx1 − x3 kX ; thus, from (5.77) we obtain kS(t)x1 − S(t)x2 kX ≤

K(1+L0 ) −ηt L0 −L kx1 − x3 kX e

.

Taking inf of the right side of this inequality with respect to x3 ∈ Y(x1 , x2 ) we conclude 0 ) that (5.64) holds, with K replaced by K(1+L L0 −L . Thus, the modified strong squeezing property holds. In conclusion, the strong squeezing property, together with the availability of a second cone invariance property with parameter L0 > L, implies the modified strong squeezing property. In this sense, the modified strong squeezing property is a weaker assumption than the original strong squeezing property.

5.6 Inertial Manifolds for Evolution Equations In this section we show how theorems 5.21 or 5.30 can be applied to construct an inertial manifold for the semiflow S generated by an autonomous evolution equation of the form (5.1) on a separable Hilbert space X with norm k · kX and scalar product h· ,·i.

5.6.1 The Evolution Problem We consider an evolution equation of the general form (5.1), that is, again, ut + Au = F(u) ,

(5.78)

where the linear operator A is densely defined, with domain dom(A) := {u ∈ X : Au ∈ X } , and generates at least a C0 -SEMIGROUP (e−tA )t ≥0 on X (see section A.3). In the next section we introduce a condition on the point spectrum of A, known as a SPECTRAL GAP CONDITION. Together with the additional assumption that F be globally bounded, i.e. F ∈ Cb (X , X ), and globally Lipschitz continuous on X , this condition will allow us to show that the semiflow S defined in X by problem (5.78) verifies the cone invariance property defined in section 5.3.1. Other, less restrictive assumptions on the point spectrum of A will assure that S also satisfies one of the two versions of the decay properties described in section 5.4 and 5.5. Consequently, theorem 5.21 (respectively, 5.30) will imply that if the operator A satisfies the spectral

5.6

Inertial Manifolds for Evolution Equations

209

gap condition, relative to the nonlinearity F, the corresponding semiflow generated by the evolution equation (5.78) admits an inertial manifold in X . Before proceeding, we remark that the conditions assumed on F, i.e. that F is globally bounded and globally Lipschitz continuous, are much stronger that those usually satisfied by the nonlinearities present in applications. For example, in the parabolic problem considered in sections 3.3, the nonlinearity F(u) = u − u3 satisfies neither condition on X (in fact, this particular function F is bounded, and Lipschitz continuous, only from bounded subsets of the subspace V ⊂ X = H into X (see proposition 3.15)). To overcome these problems, we shall suitably modify the nonlinearity outside of a bounded absorbing set, so that the nonlinearity of the modified system does satisfy the two desired conditions, and its long time dynamics coincides with that of the original one. In this sense, we take full advantage of the observation that the introduction of inertial manifolds is motivated by the desire to study the long time behavior of a system. We explain this procedure in detail in section 5.8.3, in the context of a problem concerning the Chafee-Infante equations.

5.6.2 The Spectral Gap Condition 1. Generally speaking, a spectral gap condition is a requirement on the point spectrum of A, whereby its eigenvalues should be spaced with sufficiently large gaps, so as to allow the linear part of equation (5.78), i.e. the term Au, to dominate, in a sense to be specified, the nonlinear one, i.e. the term F(u). The effect of this condition is that the corresponding semiflow S satisfies the cone invariance property. Spectral gap conditions of various type were originally introduced in the context of NavierStokes equations in two dimensions of space. Here, we introduce a specific version of the spectral gap condition, which is sufficient for the purpose of showing that the semiflows generated by problems (P) and (Hε ) admit an inertial manifold, at least if the dimension of space is n = 1 and, in (3.4), if ε is sufficiently small. For simplicity, we shall refer to this particular version as “the” spectral gap condition. Essentially, all versions of the spectral gap condition require that the point spectrum of the operator A can be divided into two parts σp (A) = σ1 ∪ σ2 ,

(5.79)

of which σ1 is finite, and such that, if Λ1 := sup{Re λ : λ ∈ σ1 } , Λ2 := inf{Re λ : λ ∈ σ2 } , the difference Λ2 − Λ1 (the “spectral gap”) is large enough, so as to satisfy an inequality of the form Λ2 − Λ1 ≥ C(Λ1ν + Λ2ν )

(5.80)

for certain nonnegative numbers C and ν related to the nonlinear term F of equation (5.78). Inequality (5.80) is usually interpreted in the sense that the spectral gap

210

5

Inertial Manifolds

should be so large as to allow the linear part of the equation to dominate the nonlinear one. To describe the particular version of the spectral gap condition we consider in the sequel, we assume that A admits a countable set of eigenvalues (λ j ) j∈N , with a corresponding orthonormal system of eigenfunctions (w j ) j∈N , spanning X (some or all of the functions w j may be complex-valued), and that the eigenvalues are ordered in such a way that Λ2 − Λ1 = Re(λN+1 − λN )

(5.81)

for some N ∈ N. Under these conditions, we give DEFINITION 5.32 Let A : X → X be as above, and assume that F ∈ Cb (X , X ) satisfies the global Lipschitz condition kF(u) − F(v)kX ≤ `F ku − vkX , for some `F independent of u and v. The operator A satisfies the CONDITION relative to F, if there is N ∈ N such that

(5.82) SPECTRAL GAP

Re(λN+1 − λN ) > 2`F .

(5.83)

That is, recalling (5.81), we assume that (5.80) holds, with ν = 0 and C = `F . 2. We now describe a condition under which the spectral gap condition (5.83) implies that the semiflow generated by equation (5.78) satisfies the cone invariance property. We fix N ∈ N, as defined in (5.81), and set X1 := span{w1 , . . . , wN } . We denote by P1 the corresponding projection of X onto X1 , and by P2 := I − P1 its complementary projection; note that A commutes with both P1 and P2 . We assume that for all u ∈ X , RehAP1 u, P1 ui ≤ Re λN kP1 uk2X ,

(5.84)

2

(5.85)

RehAP2 u, P2 ui ≥ Re λN+1 kP2 ukX .

Introducing then, for L > 0, the indefinite quadratic form HL : X → R defined by HL (u) := kP2 uk2X − L2 kP1 uk2X ,

u∈X,

(5.86)

we claim: PROPOSITION 5.33 Assume F ∈ Cb (X , X ) satisfies (5.82), that the spectral gap condition (5.83) holds, as well as conditions (5.84) and (5.85). There exist positive numbers L1 and L2 , with L1 < 1 < L2 , such that for all t0 ≥ 0, all t ≥ t0 , x and x0 ∈ X , and L ∈ [L1 , L2 ], HL (S(t)x − S(t)x0 ) ≤ HL (S(t0 )x − S(t0 )x0 ) e−2(Re λN+1 −`F )(t −t0 ) .

(5.87)

5.6

211

Inertial Manifolds for Evolution Equations

PROOF Let x, x0 ∈ X , t > 0, and set u∆ := S(t)x − S(t)x0 , F∆ := F(S(t)x) − F(S(t)x0 ) .

(5.88)

Then, recalling (5.84) and (5.85), and that, since the projections P1 and P2 are orthogonal, the identities hP2 x, P1 yi = 0 , hP2 x, P2 yi = hP2 x, yi hold for all x, y ∈ X , we can estimate (omitting as usual the reference to the variable t): 1 d HL (u∆ ) 2 dt = RehP2 u∆ , −Au∆ + F∆ i − L2 RehP1 u∆ , −Au∆ + F∆ i = − RehP2 u∆ , Au∆ i + L2 RehP1 u∆ , Au∆ i + RehP2 u∆ − L2 P1 u∆ , F∆ i ≤ − Re λN+1 kP2 u∆ k2X + L2 Re λN kP1 u∆ k2X + `F ku∆ kX kP2 u∆ − L2 P1 u∆ kX ≤ − Re λN+1 kP2 u∆ k2X + L2 Re λN kP1 u∆ k2X + 12 `F (ku∆ k2X + kP2 u∆ − L2 P1 u∆ k2X ) ≤ − Re λN+1 kP2 u∆ k2X + L2 Re λN kP1 u∆ k2X + 12 `F (kP1 u∆ k2X + kP2 u∆ k2X + kP2 u∆ k2X + L4 kP1 u∆ k2X ) . From this, we deduce the estimate d HL (u∆ ) ≤ −2γHL (u∆ ) , dt

(5.89)

where γ is any number satisfying the inequalities Re λN + 12 `F L2 + L−2 ≤ γ ≤ Re λN+1 − `F .

(5.90)

For (5.90) to hold, it is necessary that Re λN + 12 `F L2 + L−2 ≤ Re λN+1 − `F , an inequality that we can write as ψ(L) := Re λN+1 − Re λN − `F 1 + 12 L2 + L−2

≥ 0.

(5.91)

The maximum value of ψ is attained when L = 1, and the spectral gap condition (5.83) implies that ψ(1) > 0. Since ψ is concave and ψ(L) → −∞ as L → 0+ and L → +∞, it follows that ψ has two positive zeros L1 , L2 , with L1 < 1 < L2 and, as we immediately verify, L1 L2 = 1. Thus, (5.91) holds for all L ∈ [L1 , L2 ]. From (5.89) we deduce then that, with the choice γ := Re λN+1 − `F , d HL (S(t)x − S(t)x0 ) ≤ −2(Re λN+1 − `F )HL (S(t)x − S(t)x0 ) . dt Integrating this inequality on any interval [t0 ,t], with 0 ≤ t0 < t, we obtain (5.87).

212

5

Inertial Manifolds

5.6.3 The Strong Squeezing Properties In this section we show how the squeezing properties follow from the spectral gap condition (5.83). We first show that, as a consequence of proposition 5.33, the spectral gap condition implies the cone invariance property. PROPOSITION 5.34 In the same conditions of proposition 5.33, the semiflow S generated by the evolution equation (5.78) satisfies the cone invariance property in X , with parameter L, for all L ∈ [L1 , L2 ], where L1 and L2 are the positive zeros of the function ψ defined in (5.91). PROOF Let x1 , x2 ∈ X , and L ∈ [L1 , L2 ]. We have to show that if x1 − x2 ∈ CL , then S(t)x1 − S(t)x2 ∈ CL for all t ≥ 0. But, recalling (5.24), the condition x1 − x2 ∈ CL is equivalent to HL (x1 − x2 ) ≤ 0; hence, from (5.87) with t0 = 0 we deduce that HL (S(t)x1 − S(t)x2 ) ≤ 0. This means that S(t)x1 − S(t)x2 ∈ CL . We next show that if Re λN+1 is sufficiently large, the decay properties hold. PROPOSITION 5.35 In the same conditions of proposition 5.34, assume that N ∈ N is such that p η := Re λN+1 − `F 1 + L−2 > 0 ,

(5.92)

with L ∈ [L1 , L2 ]. Then S satisfies the decay property with parameters L and η. PROOF Let x1 , x2 ∈ X , and T > 0 be such that S(T )x1 − S(T )x2 ∈ / CL . We have to show that there is K > 0 such that for all t ∈ [0, T ], (5.32) holds, with η defined by (5.92) and π1 = P1 , π2 = P2 . By proposition 5.34, the semiflow S satisfies the cone invariance property; hence, by proposition 5.13, S(t)x1 − S(t)x2 ∈ / CL for all t ∈ [0, T ]. With the same definitions (5.88) of u∆ and F∆ , we have then L kP1 u∆ (t)kX ≤ kP2 u∆ (t)kX . Thus, as in the proof of proposition 5.33 (omitting again the reference to the variable t), d kP2 u∆ k2X = 2 RehP2 u∆ , F∆ − Au∆ i dt ≤ −2 Re λN+1 kP2 u∆ k2X + 2`F kP2 u∆ kX ku∆ kX ≤ −2 Re λN+1 kP2 u∆ k2X + 2`F kP2 u∆ kX (kP1 u∆ k2X + kP2 u∆ k2X )1/2 1/2 kP2 u∆ k2X ≤ −2 Re λN+1 kP2 u∆ k2X + 2`F 1 + L−2 = −2ηkP2 u∆ k2X .

5.6

Inertial Manifolds for Evolution Equations

213

From this, we conclude that for t ∈ [0, T ], kP2 (S(t)x1 − S(t)x2 )kX ≤ e−ηt kP2 (x1 − x2 )kX , and (5.32) follows, with K = 1. PROPOSITION 5.36 In the same conditions of proposition 5.34, assume that L ∈ [L1 , L2 [, and that N ∈ N is such that η1 := Re λN+1 − `F > 0 .

(5.93)

Then S satisfies the modified decay property, with parameters L and η. PROOF To prove the modified decay property, recalling (5.64) we consider x1 , x2 ∈ X and T > 0, such that P1 S(T )x1 = P1 S(T )x2 ,

(5.94)

and y ∈ Y(x1 , x2 ), i.e. such that P1 y = P1 x1 and y − x2 ∈ CL . We have to show that there is K > 0, independent of x1 and x2 , such that for all t ∈ [0, T ], kP2 (S(t)x1 − S(t)x2 )kX ≤ KkP2 (x1 − y)kX e−ηt .

(5.95)

Setting u∆ (t) := S(t)x1 − S(t)x2 as in the proof of proposition 5.33, (5.94) implies that for all L ∈ [L1 , L2 ], HL (u∆ (T )) = kP2 (u∆ (T ))k2X ≥ 0 . We claim that this implies that HL (u∆ (t)) ≥ 0 for all t ∈ [0, T ] and L ∈ [L1 , L2 ]. Indeed, if otherwise HL0 (u∆ (t 0 )) < 0 for some L0 ∈ [L1 , L2 ] and t 0 ∈ [0, T ], proposition 5.33 with t = T and t0 = t 0 would yield a contradiction to the inequality 0 ≤ HL0 (u∆ (T )). Thus, by proposition 5.33 with t0 = 0 we obtain that 0 ≤ HL (u∆ (t)) ≤ e−2(Re λN+1 −`F )t HL (x1 − x2 )

(5.96)

for all t ∈ [0, T ] and L ∈ [L1 , L2 ]. If now L ∈ [L1 , L2 [, choosing Λ ∈ ]L, L2 [ we find HΛ (u∆ (t)) = kP2 u∆ (t)k2X − Λ 2 kP1 u∆ (t)k2X = 1 − Λ 2 L2−2 kP2 u∆ (t)k2X + Λ 2 L2−2 (kP2 u∆ (t)k2X − L22 kP1 u∆ (t)k2X ) = 1 − Λ 2 L2−2 kP2 u∆ (t)k2X + Λ 2 L2−2 HL2 (u∆ (t)) ≥ 1 − Λ 2 L2−2 kP2 u∆ (t)k2X . From this, and (5.96) with L = Λ , it follows that kP2 u∆ (t)k2X ≤

L22 e−2(Re λN+1 −`F )t HΛ (x1 − x2 ) . L22 −Λ 2

(5.97)

214

5

Inertial Manifolds

Since y − x2 ∈ CL , kP2 (y − x2 )kX ≤ LkP1 (y − x2 )kX ; thus, recalling that P1 y = P1 x1 , we deduce that for all r > 0 HΛ (x1 − x2 ) = kP2 (x1 − x2 )k2X − Λ 2 kP1 (x1 − x2 )k2X ≤ (kP2 (x1 − y)kX + kP2 (y − x2 )kX )2 − Λ 2 kP1 (x1 − x2 )k2X ≤ (kP2 (x1 − y)kX + LkP1 (x2 − y)kX )2 − Λ 2 kP1 (x1 − x2 )k2X ≤ 1 + r−1 kP2 (x1 − y)k2X + ((1 + r)L2 − Λ 2 )kP1 (x1 − x2 )k2X . Choosing r = Λ 2 L−2 − 1, we obtain HΛ (x1 − x2 ) ≤

Λ2 kP2 (x1 − y)k2X Λ 2 −L2

;

replacing this into (5.97), we deduce that, for t ∈ [0, T ], kP2 u∆ (t)k2X ≤

L22 2 L2 −Λ 2

Λ2 Λ 2 −L2

e−2(Re λN+1 −`F )t kP2 (x1 − y)k2X .

This is (5.95), with K := L2Λ ((L22 − Λ 2 )(Λ 2 − L2 ))−1/2 . Since y ∈ Y(x1 , x2 ) is arbitrary, (5.64) follows, and the proof of proposition 5.36 is complete. From propositions 5.22, 5.35 and 5.36 we immediately derive COROLLARY 5.37 Assume that N is such that the spectral gap condition (5.83) holds, and Re λN+1 > `F . Then for any L ∈ [L1 , L2 [, the semiflow S satisfies the modified strong squeezing property, with parameters L and η := Re λN+1 − `F . If in addition Re λN+1 > √ `F 1 + L−2 , with L ∈ [L1 , L2 ], the semiflow√S satisfies the strong squeezing property, with parameters L and η := Re λN+1 − `F 1 + L−2 . Propositions 5.35 and 5.36 illustrate our previous remark, at the end of section 5.5.4, that the modified strong squeezing property is weaker than the strong squeezing property. Indeed, in proposition 5.36 we cannot take L = L2 , and the constant K in (5.95) is such that K → +∞ as L → L2− . On the other hand, since L < L2 , we can in fact choose L0 ∈ ]L, L2 ], and proposition 5.34 guarantees that the semiflow S does satisfy the cone invariance property, with parameter L0 > L as required in proposition 5.31. Moreover, in the modified strong squeezing property we can take smaller parameters L than for the strong squeezing property, and since (5.92) implies (5.93), the rate η1 of the exponential decay is larger.

5.6

Inertial Manifolds for Evolution Equations

215

5.6.4 Uniform Boundedness and Surjectivity We now proceed to show that the semiflow S generated by the evolution equation (5.78) satisfies the uniform boundedness property with respect to bounded subsets Φ ⊆ GL , as well as the surjectivity property with respect to arbitrary subsets Φ ⊆ GL . Here, L > 0 is arbitrary, and π1 , π2 are, of course, the projections P1 and P2 of the previous sections. As usual, we set X1 := P1 X and X2 := P2 X . Recalling that we assume the nonlinearity F : X → X to be bounded, we also set F0 := sup kF(u)k . u∈X

1.

We start with the uniform boundedness of the semiflow.

PROPOSITION 5.38 (Uniform Boundedness) Assume that there is N ∈ N such that (5.85) holds. Let Φ ⊆ GL be a bounded set of Lipschitz continuous functions ϕ : X1 → X2 . Then for all t ≥ 0 and all ϕ ∈ Φ, the set P2 S(t) graph(ϕ) is bounded in X . PROOF Given ϕ ∈ Φ and u0 ∈ graph(ϕ), fix t ≥ 0 and let u(t) := S(t)u0 . Then u solves (5.78) and, by (5.85), we can estimate d kP2 uk2X = 2 Re λN+1 hP2 u, P2 ut i = 2 Re λN+1 hP2 u, f (u) − Aui dt ≤ −2 Re λN+1 kP2 uk2X + 2F0 kP2 ukX . From this differential inequality we easily derive that for all t ≥ 0, F0 F0 e− Re λN+1 t + . kP2 u(t)kX ≤ kP2 u0 kX − Re λN+1 Re λN+1

(5.98)

(5.99)

Since u0 ∈ graph(ϕ), and Φ is a bounded set of functions, kP2 u0 kX = kϕ(P1 u0 )kX ≤ R1 for some R1 independent of u0 and ϕ. Thus, we deduce from (5.99) that kP2 u(t)kX ≤ R1 +

F0 , | Re λN+1 |

that is, kP2 S(t)u0 kX can be bounded independently of t ≥ 0 and ϕ ∈ Φ. This means that S satisfies the uniform boundedness property. We remark that (5.99) implies that for all t ≥ 0, kP2 S(t)u0 kX ≤ max kP2 u0 kX ,

K0 Re λN+1

.

216

5

Inertial Manifolds

2. We now prove two intermediate results, which we need in order to show the surjectivity property. PROPOSITION 5.39 (Backward Uniqueness) Assume S satisfies the cone invariance property with parameter L ∈ [L1 , L2 ]. Assume that (5.84) holds for some N ∈ N, and let u0 , v0 ∈ X and T > 0 be such that P1 S(T )u0 = P1 S(T )v0 and HL (u0 − v0 ) ≤ 0. Then u0 = v0 . PROOF Let, as before, u∆ (t) := S(t)u0 − S(t)v0 , F∆ (t) := F(S(t)u0 ) − F(S(t)v0 ) . The condition HL (u0 − v0 ) ≤ 0 implies that u0 − v0 ∈ CL ; then, by the cone invariance property, u∆ (t) ∈ CL for all t ≥ 0, i.e. kP2 u∆ (t)kX ≤ LkP1 u∆ (t)kX .

(5.100)

Recalling (5.84), we can therefore estimate (omitting as usual the argument t) d kP1 u∆ k2X = 2 Re λN hP1 u∆ , F∆ − Au∆ i dt ≥ −2 Re λN kP1 u∆ k2X − 2`F ku∆ kX kP1 u∆ kX p ≥ −2 Re λN kP1 u∆ k2X − 2`F 1 + L2 kP1 u∆ k2X . Integrating this inequality on [t, T ], 0 ≤ t < T , we find √ Re λN +`F 1+L2 (T −t) kP1 u∆ (t)kX ≤ e kP1 u∆ (T )kX ; and since P1 u∆ (T ) = P1 (S(T )u0 − S(T )v0 ) = 0 , we conclude that P1 u∆ (t) = 0 for all t ∈ [0, T ]. Together with (5.100), this implies that u∆ (t) ≡ 0, i.e. that S(t)u0 ≡ S(t)v0 in [0, T ]. PROPOSITION 5.40 (Coercivity) Under the same conditions of proposition 5.39, the semiflow S is COERCIVE, i.e. for all x ∈ X and t ≥ 0, kP1 S(t)xkX → +∞ as kP1 xkX → +∞ .

(5.101)

PROOF Proceeding as in (5.98), but considering P1 u instead of P2 u, we obtain the estimate d kP1 uk2X ≥ −2 Re λN kP1 uk2X − 2F0 kP1 ukX , dt

5.6

217

Inertial Manifolds for Evolution Equations

from which we deduce that for all t ≥ 0, kP1 u(t)kX ≥ kP1 u0 kX e− Re λN t −

F0 (1 − e− Re λN t ) . Re λN

Keeping in mind that t > 0 is kept fixed, this inequality implies (5.101). 3. We can now show that S satisfies the surjectivity property, with respect to all of GL . PROPOSITION 5.41 (Surjectivity) Assume S satisfies the cone invariance property with parameter L ∈ [L1 , L2 ]. Then for all ϕ ∈ GL and t ≥ 0, P1 S(t) graph(ϕ) = P1 X . PROOF The inclusion P1 S(t) graph(ϕ) ⊆ P1 X is obvious. For the inverse inclusion, fix ϕ ∈ GL and t > 0, and set Φ := P1 S(t) graph(ϕ) ⊆ P1 X . We define a function h : X1 → X1 by h(ζ ) := P1 S(t)(ζ + ϕ(ζ )) ,

ζ ∈ X1 .

(5.102)

Since ϕ, S(t) and P1 are continuous, h is also continuous. We now show that h is also injective. Indeed, let ζ0 and ζ1 ∈ X1 be such that h(ζ0 ) = h(ζ1 ), and set u0 = ζ0 + ϕ(ζ0 ), u1 = ζ1 + ϕ(ζ1 ). Then, u0 and u1 ∈ graph(ϕ). Since ϕ ∈ GL , kP2 (u0 − u1 )kX = kϕ(ζ0 ) − ϕ(ζ1 )kX ≤ Lkζ0 − ζ1 kX = LkP1 (u0 − u1 )kX ; thus, recalling (5.86), HL (u0 − u1 ) ≤ 0. By (5.102), the condition h(ζ0 ) = h(ζ1 ) means that P1 S(t)u0 = P1 S(t)u1 ; hence, by proposition 5.39, u0 = u1 . Therefore, ζ0 = ζ1 , which means that h is injective. We can then define the inverse function h−1 : Φ → X1 . We now show that h−1 is also continuous. Arguing by contradiction, assume that there is ξ ∈ Φ such that h−1 is not continuous at ξ . We can then find a number ε0 > 0, and a sequence (ξk )k∈N ⊂ Φ, such that ξk → ξ as k → ∞, but for all k ∈ N kh−1 (ξk ) − h−1 (ξ )kX ≥ ε0 .

(5.103)

Let ζ := h−1 (ξ ), ζk := h−1 (ξk ), zk := ζk + ϕ(ζk ), and z := ζ + ϕ(ζ ). Then ξk = h(ζk ) = P1 S(t)zk , and ξ = h(ζ ) = P1 S(t)z. The coercivity property implies that the sequence (ζk )k∈N is bounded in X1 : in fact, if otherwise kζk kX = kP1 zk kX → +∞ as k → +∞, then by (5.101) also kP1 S(t)zk kX = kξk kX → +∞, contradicting the fact that ξk → ξ . Since P1 X is finite dimensional, there is a subsequence (ζkm )m∈N ,

218

5

Inertial Manifolds

converging to some limit ζ ∈ P1 X . The continuity of h implies that h(ζkm ) → h(ζ ); but since h(ζkm ) = ξkm , we conclude that h(ζ ) = ξ , i.e. that ζ = h−1 (ξ ). Thus, h−1 (ξkm ) → h−1 (ξ ), contradicting (5.103). It follows that h−1 is continuous. By proposition 5.40, h is coercive on P1 X ; since this space is finite dimensional, h is a sequentially continuous mapping. Then, by theorem A.44, h is a homeomorphism from X1 into itself. This concludes the proof of proposition 5.41. 4. In conclusion, we have seen that if operator A of the evolution equation (5.78) satisfies the spectral gap condition (5.83), with N sufficiently large so that (5.93) or the stronger inequality (5.92) holds, then the corresponding semiflow S admits an inertial manifold M ⊆ X , of the form (5.16). Since we also know that S admits a compact, positively invariant absorbing set, we conclude that S admits a compact inertial manifold. We summarize this result in THEOREM 5.42 Let A be a densely defined operator generating a C0 -semigroup on a separable Hilbert space X . Let F ∈ Cb (X , X ) satisfy the Lipschitz condition (5.82). Assume A satisfies the spectral gap condition (5.83), with N so large that Re λN+1 > `F .

(5.104)

Then the semiflow S generated by the autonomous evolution equation (5.78) admits an inertial manifold M in X , such that dimF (M) = dim X1 . PROOF By corollary 5.37, S satisfies at least the modified strong squeezing property, with parameters L ∈ [L1 , L2 ] and η = Re λN+1 − `F . By propositions 5.38 and 5.41, S also satisfies the uniform boundedness and surjectivity properties, with respect to any bounded subset Φ ⊆ GL . The existence of an inertial manifold M follows then from theorem 5.30. Since M is the graph of a Lipschitz continuous function over X1 , its fractal dimension is the same as that of X1 (i.e., N). √ REMARK 5.43 1. If the stronger condition Re λN+1 > `F 1 + L−2 holds, the existence of M would also follow from theorem 5.21. 2. If the point spectrum of A is such that Re λN ≥ 0 for all N ∈ N, then condition (5.104) in theorem 5.42 is redundant, since it follows from the spectral gap condition (5.83). Indeed, in this case we have that Re λN+1 > Re λN + 2`F > `F .

5.7 Applications In this section we briefly describe how the spectral gap condition can be formulated for parabolic and hyperbolic evolution problems of the type (P) and (Hε ). In

5.7

219

Applications

the next section, we shall see that the spectral gap condition is actually satisfied for the specific situation when, in these problems, the space dimension is n = 1 and, for problem (Hε ), ε is sufficiently small.

5.7.1 Semilinear Heat Equations We consider a semilinear heat equation of the form ut − ∆u = f (x) − g(u) ,

(5.105)

in the space X := L2 (Ω ). As in (3.1), we subject u to homogeneous Dirichlet boundary conditions, but we assume that f : X → X is globally bounded and globally Lipschitz continuous, with Lipschitz constant `. In this case, equation (5.105) is directly in the form (5.78), with A = −∆, dom(A) = H2 (Ω ) ∩ H10 (Ω ), and F(u) := f − g(u). By theorem A.79, A generates an analytic semigroup of operators in X . Moreover, as seen in section 3.2.2, A admits a complete system of eigenfunctions {w j } j∈N , corresponding to all real eigenvalues 0 < λ1 ≤ λ2 ≤ · · · → +∞ (see theorem A.76). Then, each u ∈ X admits the uniquely determined Fourier series expansion (3.21), that is, ∞

u=

∑ hu, w j iw j

(5.106)

j=1

(the series converging in X ), and Parseval’s formula (3.22), i.e. ∞

kuk2X =

∑ hu, w j i2 ,

(5.107)

j=1

holds for all u ∈ X . Then, ∞

dom(A) = {u ∈ X :

∑ λ j hu, w j iw j

converges in X } .

j=1

Let u ∈ dom(A). Then, hAu, ui is real and nonnegative, and, by (5.106), + * * ∞

hAu, ui =

∞

∑ hu, w j iAw j , ∑ hu, wk iwk

j=1

k=1

∞

=

+

∞

∑ hu, w j iλ j w j , ∑ hu, wk iwk

j=1

k=1

∞

=

∑ λ j hu, w j i2 ,

(5.108)

j=0

the last step being a consequence of the orthogonality of the eigenfunctions w j . As in section 3.2.3, for N ∈ N>0 we set X1 := span{w1 , . . . , wN } , and define the corresponding projection P1 : X → X1 by (3.27). Setting as usual P2 := I − P1 , we see that (5.108) implies that A satisfies conditions (5.84) and (5.85).

220

5

Inertial Manifolds

In fact, if v := P1 u, then hv, w j i = 0 for all j ≥ N + 1, so (5.108) and (5.107) imply that N

N

hAv, vi =

∑ λ j hv, w j i2 ≤ λN ∑ hv, w j i2 = λN kvk2X .

(5.109)

j=1

j=1

Likewise, if w := P2 u, ∞

hAw, wi =

∑

λ j hw, w j i2 ≥ λN+1 kwk2X .

(5.110)

j=N+1

It follows that proposition 5.33 can be applied, and the semiflow S generated by equation (5.105) satisfies the cone invariance property, with parameter L, provided that N is so large that the spectral gap condition λN+1 − λN ≥ 2`

(5.111)

holds. As we observed in remark 5.43, condition (5.104) holds, since λN ≥ 0 for all N ∈ N>0 . In conclusion, if N is so large that (5.111) holds, the semiflow S generated by the parabolic PDE (5.105) admits an inertial manifold in X . In section 5.8, we shall see that the spectral gap condition (5.111) can certainly be satisfied if the space dimension is n = 1.

5.7.2 Semilinear Wave Equations We consider a semilinear wave equation of the form εutt + ut − ∆u = f (x) − g(u) ,

ε > 0.

(5.112)

As in (3.4), we subject u to homogeneous Dirichlet boundary conditions, but, as in (5.105), we assume g : L2 (Ω ) → L2 (Ω ) to be globally bounded and globally Lipschitz continuous, with Lipschitz constant `. Equation (5.112) is formally equivalent to a first order system of the form (3.14), i.e. Ut + AU = F(U) for U = (u, v) ∈ X := H1 (Ω ) × L2 (Ω ), where ! 0 − ε1 A= , F(U) = 1 −∆ ε

(5.113)

0 f − g(u)

! ,

(5.114)

and dom(A) = H2 (Ω ) ∩ H10 (Ω ) × H1 (Ω ) (we identify for convenience pairs (u, v) of functions with the corresponding column vector (u, v)> ). The operator A defined in (5.114) generates a C0 -semigroup in X (see theorem A.53). However, this semigroup is not analytic, and A is neither self-adjoint nor positive. In particular, conditions (5.84) and (5.85) do not hold. Still, by theorem

5.7

221

Applications

3.20 we know that (5.113) generates a semiflow S in X (for an alternative proof directly concerning system (5.113), see e.g. Sell and You, [SY02, sct. 5.2]). Our goal is to show that, if ε is sufficiently small, A satisfies the spectral gap condition, and, consequently, S admits an inertial manifold in X . To this end, we will consider in X an equivalent norm, introduced by Mora in [Mor87]. √ In order to conform to Mora’s presentation, we introduce the time rescaling t 7→ εt; that is, we consider the new unknown √ w(x,t) := u(x, εt) , which solves the equation wtt + √1ε wt − ∆w = f (x) − g(w) . Setting then α :=

1 √ , 2 ε

and renaming w, we finally consider the equation utt + 2αut − ∆u = f (x) − g(u) .

This equation is formally equivalent to system (5.113), where now ! 0 −1 . A= −∆ 2α

(5.115)

Moreover, we can now consider in X the standard graph norm, i.e. k(u, v)k2X = k∇uk2 + kvk2 ,

(u, v) ∈ X .

(5.116)

To determine the spectrum of A, we observe that the eigenvalue equation AU = µU ,

U = (u, v) ∈ X ,

is equivalent to the system (

v = −µu −∆u + 2αv = µv .

(5.117)

Thus, u must satisfy the conditions u ∈ H10 (Ω ) ,

− ∆u = (2α µ − µ 2 )u .

It follows that 2α µ − µ 2 must be an eigenvalue of −∆, with respect to homogeneous Dirichlet boundary conditions; that is, if (λ j ) j∈N denotes the sequence of these eigenvalues, we must have 2α µ − µ 2 = λ j ,

j ≥ 1.

For each j ≥ 1, this equation has the two complex solutions q := α ± µ± α2 − λ j ; j

(5.118)

(5.119)

222

5

Inertial Manifolds

thus, A does have a countable set of eigenvalues. Since min j≥1 λ j = λ1 > 0, we see that if ε is so large that α2 =

1 4ε

< λ1 ,

then no eigenvalue of A is real. Moreover, since in this case all eigenvalues have the same real part α, it is clearly impossible to find any decomposition of the point spectrum σp (A), as in (5.79), such that the spectral gap condition (5.83) holds. Hence, the existence of an inertial manifold for equation (5.112) is, in general, open. Indeed, as we have mentioned, in chapter 7 we present an explicit non-existence result for a similar problem, which holds precisely when ε is sufficiently large. On the other hand, when ε is small (i.e., when α is large), some eigenvalues of A will be real, positive and distinct, so that it makes sense to investigate whether the spectral gap condition may hold. In fact, we have the following result. THEOREM 5.44 Let ` be the Lipschitz constant of f . Assume there is m ∈ N such that λm+1 − λm > 4` ,

(5.120)

and let N be the smallest integer such that (5.120) holds. Assume that α is so large that α 2 > λN+1 .

(5.121)

There is then an equivalent norm in X , such that the operator A defined in (5.115) satisfies the spectral gap condition relative to F. As a consequence, the semiflow S generated by system (5.113) admits an inertial manifold in X , whose fractal dimension does not exceed N. PROOF 1. We set N0 := { j ∈ N : λ j ≤ α 2 } , N1 := { j ∈ N : λ j > α 2 } . Then, N0 is not empty, because λN+1 ∈ N0 by (5.121). The eigenvalues µ ± j of A are ± ± such that µ j ∈ R if j ∈ N0 , while µ j ∈ C \ R if j ∈ N1 ; in the latter case, Re µ ± j =α.

(5.122)

We decompose the point spectrum of A as in (5.79), with σ1 := {µ1− , . . . , µN− } ; thus, Λ1 = µN− = α − µk+ ∈ R, and

p

α 2 − λN . To determine Λ2 , observe first that if k ∈ N0 , then

µk+ = α +

p p − α 2 − λk ≥ α − α 2 − λN+1 = µN+1 ;

(5.123)

5.7

223

Applications

also, if k ∈ N0 and k ≥ N + 2, then µk− ∈ R, and p p − µk− = α − α 2 − λk ≥ α − α 2 − λN+1 = µN+1 .

(5.124)

− Together with (5.123) and (5.122), (5.124) implies that Λ2 = µN+1 ; consequently,

Λ2 − Λ1 =

p

α 2 − λN −

p α 2 − λN+1 .

(5.125)

Corresponding to the eigenvalues (5.119), A has eigenfunctions U j± , which are complex-valued if j ∈ N1 . By (5.117), these eigenfunctions have the form U j± = (u j , −µ ± j u j) ,

with − ∆u j = λ j u j .

(5.126)

As we immediately verify, the system (U j± ) j∈N is linearly independent and complete in X ; that is, span{U j± : j ∈ N} = X (closure in X ). Thus, for N ∈ N>0 we can set X1 := span{U1− , . . . ,UN− } .

(5.127)

This subspace of X is finite dimensional, but is not orthogonal to the subspace − − + + X2 := span{U1+ , . . . ,UN+ ,UN+1 ,UN+1 ,UN+2 ,UN+2 ,···}

(5.128)

spanned by the other eigenfunctions of A. To see this, it is sufficient to note that, because of (5.126), for 1 ≤ j ≤ N − + 2 2 + hU j− ,U j+ iX = k∇u j k2 + h−µ − j u j , −µ j u¯ j i = λ j ku j k + µ j µ j ku j k

= 2λ j 6= 0 , + having recalled (5.116), and that ku j k2 = 1, k∇u j k2 = λ j (by (3.24)), and µ − j µ j = λ j. 2. To overcome this difficulty, we define a new norm in X , equivalent to the norm (5.116), with the purpose of making X1 and X2 orthogonal with respect to this new norm. Then, we will see that, as a consequence of (5.120) and (5.121), the operator A satisfies the spectral gap condition relative to F, expressed in term of the new norm. Adapting Mora’s presentation of [Mor87], we further decompose σ2 into the sets

σ21 := {µ1+ , . . . , µN+ } , σ2∞ := (µ ± j ) j≥N+1 and define corresponding subspaces of X X21 := span{U1+ , . . . ,UN+ } , X2∞ := span{U j± , j ≥ N + 1} ; we also set X11 := span{U j± : j ≤ N} = X1 ⊕ X21 .

224

5

Inertial Manifolds

Let U = (u, v) ∈ X . We immediately note that if U ∈ X1 , or U ∈ X21 , then, as a consequence of (5.126), ( −µ − if U ∈ X1 , j hu, u j i hv, u j i = −µ + hu, u i if U ∈ X21 , j j for 1 ≤ j ≤ N. On the other hand, if U ∈ X11 or U ∈ X2∞ , there is no special relations between the Fourier coefficients of the components of U, and we have the standard decompositions  N   if U ∈ X11 ,  U = ∑ (hu, u j iu j , hv, u j iu j ) j=1

∞

  U =

∑ (hu, u j iu j , hv, u j iu j )

if U ∈ X2∞ .

j=N+1

To see this, let e.g. U ∈ X2∞ . Then ∞

U=

∑

(ξ j−U j− + ξ j+U j+ ) ,

(5.129)

j=N+1

for suitable sequences (ξ j± ) j≥N+1 ⊂ C (note that the second component of U j± is / R if j ∈ N1 ). Recalling (5.126), (5.129) can be complex-valued, because µ ± j ∈ written as ∞

U=

∑

+ + ((ξ j− + ξ j+ )u j , (ξ j− µ − j + ξ j µ j )u j ) ;

j=N+1

thus, we deduce that the Fourier coefficients of u and v must satisfy the system ( hu, u j i = ξ j− + ξ j+ , (5.130) − + + hv, u j i = −(µ − j ξj + µj ξj ). Since the determinant of the matrix 1 µ− j

1 µ+ j

!

p − 2 equals µ + j − µ j = 2 α − λ j 6= 0, system (5.130) uniquely determines each pair − + (ξ j , ξ j ), and vice versa. Finally, we remark that, in analogy with (5.109) and (5.110), we have the implications (u, v) ∈ X11 (u, v) ∈ X2∞

=⇒ =⇒

k∇uk2 ≤ λN kuk2 , 2

(5.131) 2

k∇uk ≥ λN+1 kuk ,

which follow from the Fourier series expansion of u in L2 (Ω ).

(5.132)

5.7

225

Applications

3. We now define a function Φ : X11 × X11 → R, by Φ(U,V ) := α 2 hu, yi ¯ − h∇u, ∇yi ¯ + hαu + v, α y¯ + z¯i , for U = (u, v) and V = (y, z) ∈ X11 ; the bar denotes complex conjugation. Because of (5.131), and recalling that λN < λN+1 < α 2 =

1 4ε

,

we have that for all U ∈ X11 , Φ(U,U) = α 2 kuk2 − k∇uk2 + kαu + vk2 ≥ (α 2 − λN )kuk2 ≥ 0 .

(5.133)

Thus, Φ is positive definite, and therefore defines a scalar product on X11 ; we denote by k · k11 the corresponding norm. We also denote by k · k2∞ the norm induced in the orthogonal space X2∞ by the standard norm (5.116); an analogous notation denotes the corresponding scalar product in X2∞ . We are now ready to redefine the scalar product and the norm in X ; to avoid introducing more symbols than necessary, for the remainder of this section we agree to keep the notations h · , · iX and k · kX to denote these new scalar product and norm. We fix R > 0 such that R≥

α2 , λN+1

(5.134)

and define the new scalar product in X by hU,V iX := Φ(P11U, P11V ) + RhP2∞U, P2∞V i2∞ ,

(5.135)

where P11 and P2∞ are the projections from X into, respectively, X11 and X2∞ , defined by means of the Fourier series expansion of U ∈ X (for instance, P2∞U is defined by (5.129)). With some abuse of notation, we shall abbreviate (5.135) by hU,V iX := Φ(U,V ) + RhU,V i2∞ .

(5.136)

It is then easy to see that, as a consequence of (5.131), (5.136) does define an equivalent norm in X . We now show that the spaces X1 and X2 , defined in (5.127) and (5.128), are now orthogonal with respect to this new norm. PROPOSITION 5.45 Let X be endowed with the scalar product (5.136). Let j, k ≥ 1. Then: hU j± ,Uk± iX = hU j± ,Uk∓ iX = 0 if j 6= k ; ± 2

(5.137)

2

kU j k11 = 2(α − λ j ) if 1 ≤ j ≤ N ; ( p 2 −λ 2α α ± α N+1 kU j± k22∞ = 2λ j

(5.138) if j ≥ N + 1, j ∈ N0 , if j ≥ N + 1, j ∈ N1 .

(5.139)

226

5

Inertial Manifolds

In particular, X1 ⊥ X21 and X21 ⊥ X2∞ . Therefore, since X2 = X21 ⊕ X2∞ , X1 ⊥ X2 .

(5.140)

PROOF All the identities in (5.137) are an immediate consequence of (5.126) and the orthogonality of the system (u j ) j≥1 , both in L2 (Ω ) and H10 (Ω ). In particular, this implies the orthogonality of X11 and X2∞ . To see that X1 ⊥ X21 , we still have to show that, in addition to (5.137) for 1 ≤ j, k ≤ N, also hU j− ,U j+ iX = 0 ,

1≤ j≤N.

(5.141)

To see this, recalling (5.133) and (5.126) we compute hU j− ,U j+ iX = Φ(U j− ,U j+ ) + = α 2 ku j k2 − k∇u j k2 + h(α − µ − j )u j , (α − µ j )u¯ j i − − + 2 2 = 2α 2 ku j k2 − λ j ku j k2 − α(µ + j + µ j )ku j k + µ j µ j ku j k . − + − By (5.118), µ + j + µ j = 2α, and µ j µ j = λ j ; hence, (5.141) follows. Identity (5.138) is a consequence of (5.133), by which we have 2 kU j± k211 = Φ U j± ,U j± = α 2 ku j k2 − k∇u j k2 + k(α − µ ± j )u j k q 2 2 2 2 = (α − λ j )ku j k + ∓ α − λ j ku j k2 ,

recalling that ku j k2 = 1. As for (5.139), acting likewise we find that 2 kU j± k22∞ = λ j + |µ ± | ku j k2 . j If j ∈ N0 , µ ± j =α±

p α 2 − λ j ∈ R, and

q q 2 2 λ j + |µ j | = λ j + α + α − λ j ± 2α α − λ j = 2α α ± α − λ j , ± 2

2

2

p 2 while if j ∈ N1 , µ ± j = α ± i λ j − α ∈ C \ R, and 2 2 2 λ j + |µ ± j | = λ j + α + λ j − α = 2λ j .

This concludes the proof of proposition 5.45. 4. Having thus established the desired orthogonal decomposition (5.140), we call P1 : X → X1 and P2 : X → X2 the corresponding orthogonal projections, and proceed to show that A, P1 and P2 satisfy inequalities (5.84) and (5.85). We actually show that ∀U ∈ X1 :

hAU,UiX = hU, AUiX ≤ µN− kUk2X ,

(5.142)

5.7

227

Applications

− RehAU,UiX = RehU, AUiX ≥ µN+1 kUk2X .

∀U ∈ X2 :

(5.143)

In fact, these inequalities are immediate. To show (5.142), let U ∈ X1 . Then, N

U=

∑ γ jU j− ,

γ j := hu, u j i ;

j=1

thus, AU =

N

N

j=1

j=1

∑ γ j AU j− = ∑ γ j µ −j U j− ,

and, therefore, by (5.137) N

hAU,UiX = Φ(AU,U) = Φ

−

N

−

! −

∑ γ j µ j U j , ∑ γkUk

j=1

k=1

N

= ∑ γi2 µi− kUi− k211 = Φ(U, AU) = hU, AUiX . i=1

This shows the first half of (5.142). The second half follows from the inequality − µ− j ≤ µN , 1 ≤ j ≤ N, and recalling that if U ∈ X1 , again by (5.137) ! kUk2X = kUk211 = Φ

N

N

j=1

k=1

∑ γ jU j− , ∑ γkUk−

N

= ∑ γi2 kUi− k211 .

(5.144)

i=1

To show (5.143), let U ∈ X2 . Then, N

U=

∞

∑ γ jU j+ +

j=1

∑

(ξ j−U j− + ξ j+U j+ ) =: U21 +U2∞ ,

j=N+1

with γ j := hu, u j i ∈ R and ξ j± ∈ C, as in (5.129). Again by (5.137), hAU,UiX = Φ(AU21 ,U21 ) + hAU2∞ ,U2∞ i2∞ ∞

N

= ∑ γi2 µi+ kUi+ k211 + i=1

∑

(|γi− |2 µi− kUi− k22∞ + |γi+ |2 µi+ kUi+ k22∞ ) ,

i=N+1

hU, AUiX = Φ(U21 , AU21 ) + hU2∞ , AU2∞ i2∞ N

∞

= ∑ γi2 µi+ kUi+ k211 + i=1

∑

(|γi− |2 µi− kUi− k22∞ + |γi+ |2 µi+ kUi+ k22∞ ) .

i=N+1

Thus, RehAU,UiX = RehU, AUiX . The second half of (5.143) follows then from (5.122), and from the obvious inequalities q p = α + µ+ α 2 − λ j > α − α 2 − λN+1 if j ∈ N0 , j ≤ N , j

228

5 µ− j =α−

q

Inertial Manifolds

α2 − λ j ≥ α −

p α 2 − λN+1

if j ∈ N0 , j ≥ N + 1 .

Indeed, these inequalities imply that − RehAU,UiX ≥ µN+1 −

N

∞

i=1 2

i=N+1

∑ γi2 kUi+ k211 + ∑

|γi− |2 kUi− k22∞ + |γi+ |2 kUi+ k22∞

= µN+1 kUkX , the last identity being established as in (5.144). 5. Our last step is to estimate the Lipschitz constant of F in (5.113); recall that F(U) := (0, f − g(u)). To this end, we first remark that the orthogonal projections P11 and P2∞ introduced in (5.135) induce corresponding orthogonal projections p11 and p2∞ in H10 (Ω ) and L2 (Ω ), naturally defined as follows. If U = (u, v) ∈ X and Ui j := (ui j , vi j ) = Pi jU, i j = 11 or i j = 2∞, then pi j u := ui j , pi j v := vi j . Thus, from (5.133) and (5.132) it follows that, for any U = (u, v) ∈ X , kUk2X = kP11Uk211 + RkP2∞Uk22∞ ≥ (α 2 − λN )kp11 uk2 + Rk∇p2∞ uk2 + kp2∞ vk2 ≥ (α 2 − λN )kp11 uk2 + RλN+1 kp2∞ uk2 .

(5.145)

From (5.134) we deduce that RλN+1 ≥ α 2 − λN ; hence, we obtain from (5.145) that kUk2X ≥ (α 2 − λN ) kp11 uk2 + kp2∞ uk2 = (α 2 − λN )kuk2 .

(5.146)

Given then U = (u, v) and V = (y, z) ∈ X , we compute kF(U) − F(V )k2X = kP11 (F(U) − F(V ))k2X + kP2∞ (F(U) − F(V ))k2X = kp11 ( f (u) − f (v))k2 + kp2∞ ( f (u) − f (v))k2 = k f (u) − f (v)k2 ≤ `2 ku − vk2 ≤

`2 kU −V k2X , α 2 − λN

the last step following from (5.146). Thus, ` `F ≤ p . 2 α − λN

(5.147)

5.8

229

Semilinear Evolution Equations in One Space Dimension

6. We can now conclude the proof of theorem 5.44. From (5.125) and (5.147) we deduce that the operator A defined by (5.115) satisfies the spectral gap condition (5.83), relative to F, if there exists N ∈ N such that λN+1 < α 2 and p p 2` α 2 − λN − α 2 − λN+1 > p . α 2 − λN Multiplying by inequality

(5.148)

p α 2 − λN , and then squaring, we see that (5.148) is equivalent to the (α 2 − λN )(λN+1 − λN − 4`) + 4`2 > 0 ,

which in turn is implied by (5.120) (with the lowest value m = N). As observed in remark 5.43, condition (5.104) also holds, since Re µN ≥ 0 for all N ∈ N>0 . Thus, theorem 5.42 can be applied, and theorem 5.44 follows. In conclusion, let N be the lowest possible value of the integers m such that condition (5.120) holds. Then, if α is so large that (5.121) holds, then the semiflow S generated by the semilinear damped wave equation (5.112) admits an inertial manifold in X . In section 5.8.5 we shall see that conditions (5.120) and (5.121) can certainly be satisfied if, again, the dimension of space is n = 1.

5.8 Semilinear Evolution Equations in One Space Dimension In this section we show that, when the dimension of space is n = 1, theorem 5.42 can be applied to problems (P) and (Hε ) (the latter, at least if ε is sufficiently small). As a consequence, in these cases we can deduce the existence of an inertial manifold for the corresponding semiflows. In the parabolic problem, the estimates we obtain will also allow us to slightly improve our results of section 3.3.1, concerning the absorbing sets of the related semiflow.

5.8.1 The Parabolic Problem We consider the Chafee-Infante reaction-diffusion equation (3.1) in one space dimension, i.e. the IBVP on ]0, +∞[ × ]0, π[  3   ut − uxx + k(u − u) = f (x) (5.149) u(0, x) = u0 (x)   u(t, 0) = u(t, π) = 0 , where k > 0. From the results of chapter 3, we know that problem (5.149) generates a semiflow S, both in X = L2 (0, π) and in V = H10 (0, π). We also know that S admits a bounded, positively invariant absorbing set in both these spaces. In particular, since

230

5

Inertial Manifolds

H10 (0, π) ,→ C([0, π]), S has a bounded absorbing set in L∞ (0, π). S also admits a compact attractor A ⊆ X , which is contained and bounded in V. When f ≡ 0, the structure of this attractor is known in large detail, at least when k is not the square of an integer; indeed, the following results hold (see e.g. Jolly, [Jol89]). THEOREM 5.46 − Assume that m2 < k < (m + 1)2 , m ∈ N. Then, there are m pairs (ϕ + j , ϕ j ), j = 0, . . . , m − 1, of nontrivial equilibria for (5.149), having the following properties: − 1. Each ϕ + j and ϕ j , j = 0, . . . , m − 1, is an hyperbolic equilibrium (in the sense of definition 2.21), is connected to the origin (i.e., to the zero solution) by a heteroclinic orbit (see definition 2.24), and has exactly j simple zeroes in ]0, π[.

2. If 0 ≤ j ≤ i ≤ m − 1, there also exist connecting heteroclinic orbits from ϕ ± j to ± ϕi . 3. The global attractor is the union of the unstable manifolds of these equilibria. The result on the zeroes of the equilibria ϕ ± j is shown in Chafee-Infante, [CI74]. A proof of the other results can be found e.g. in Henry, [Hen85], where additional results on the dimension of the global attractor are also given. The existence of an inertial manifold for problem (5.149) is also known: see e.g. Jolly, [Jol89], where an inertial manifold of the form (5.16) is constructed, resorting to a local extension of the unstable manifolds along suitable stable directions. Consequently, S also has an exponential attractor E, which is obtained as the intersection of the inertial manifold and a compact absorbing, positively invariant ball B (whose existence we mentioned above). In the sequel, we will assume that f ≡ 0 for simplicity, and show how the inertial manifold for S can be obtained as an application of theorem 5.42. Thus, we have to show that the semiflow S satisfies the spectral gap condition (5.83), with respect to the nonlinearity F(u) := k(u − u3 ). This requires that we overcome the problem that, as we have already remarked, F is not Lipschitz continuous from X into X , but only from bounded sets of X1 into X , as seen in proposition 3.15. (Actually, the proof of that proposition shows that F is Lipschitz continuous from bounded sets of L∞ (0, π) into X .) We shall take care of this problem by suitably “adjusting” F outside a fixed absorbing set B in L∞ (0, π), so as to make F globally bounded and Lipschitz continuous. Since both the attractor and the inertial manifold are contained in B, and B is absorbing, this adjustment of F will not affect the long-time dynamics of the system.

5.8.2 Absorbing Sets In this section we prove the existence of bounded, positively invariant absorbing sets for S in the space L∞ (0, π), without resorting to the imbedding of V into this space. Rather, as in Eden, Foias, Nicolaenko and Temam, [EFNT94], we establish

5.8

Semilinear Evolution Equations in One Space Dimension

231

estimates on the norms ku(·,t)kL2p (0,π) , 1 ≤ p ≤ ∞, which allow us not only to deduce the existence of this absorbing set, but also to obtain a specific control on the Lipschitz constant ` of the nonlinearity F. We will use this new result to show that, when we introduce the stated modification of F, the modified equation satisfies the spectral gap condition (5.83). In the sequel, we denote as usual by | · | p the norm in L p (0, π), for 1 ≤ p ≤ +∞, and, for a > 0, we set B p (a) := {u ∈ H10 (0, π) : |u| p ≤ a} . We claim: LEMMA 5.47 For all a > 0, B∞ (a) =

\

B2p (π 1/2p a) .

p≥1

PROOF The inclusion B∞ (a) ⊆

\

B2p (π 1/2p a)

p≥1

follows from the estimate |u|2p ≤ π 1/2p |u|∞ . As for the opposite inclusion, since |u|∞ = lim |u|2p , p→∞

lim π 1/2p = 1 ,

p→∞

(5.150)

given arbitrary ε > 0 there is p0 ≥ 1 such that for all p ≥ p0 , |u|∞ ≤ |u|2p + ε ,

π 1/2p a ≤ a + ε .

Consequently, for u ∈ B2p (π 1/2p a), p ≥ p0 , |u|∞ ≤ π 1/2p a + ε ≤ a + 2ε . Letting ε → 0, we deduce that |u|∞ ≤ a, that is, u ∈ B∞ (a). We proceed then to prove PROPOSITION 5.48 If a ≥ π 1/2p , 1 ≤ p < +∞, B2p (a) is positively invariant with respect to S. The same is true for B∞ (a) if a ≥ 1. PROOF Let u be a solution of (5.149) and, for 1 ≤ r ≤ ∞, set vr (t) := |u(t, ·)|r .

232

5

Inertial Manifolds

Our first step is an estimate of v2p (t); for convenience, we sometimes omit the dependence of the functions from some or all of their variables. For each p ∈ [1, +∞], we compute Z Z π 1 d 2p 1 d π v2p (t) = |u(t, x)|2p dx = p u2p−1 ut dx 2 dt 2 dt 0 0 Z π

=p

u2p−1 (k(u − u3 ) + uxx ) dx

0

Z π

= pk

u2p dx − pk

0

Z π

u2p+2 dx + p

0

0 2p−1

= pk(v2p )2p − pk(v2p+2 )2p+2 + p[u − p(2p − 1)

Z π

Z π

u2p−1 uxx dx

ux ]x=π x=0

u2p−2 (ux )2 dx .

0

Because of the boundary conditions u(t, 0) = u(t, π) = 0, the terms within the square brackets equal zero; hence, neglecting the last term, which is negative, we obtain the inequality d (v2p )2p ≤ 2pk(v2p )2p − 2pk(v2p+2 )2p+2 . dt

(5.151)

By Hölder’s inequality, 2p

(v2p (t))

Z π

=

|u(t, x)|2p dx ≤ π 1/(p+1)

0 1/(p+1)

=π

Z

π

|u(t, x)|2p+2 dx

p/(p+1)

0 2p

(v2p+2 (t)) ;

thus, v2p+2 (t) ≥ π −1/2p(p+1) v2p (t) , and from (5.151) we obtain the estimate d (v2p )2p ≤ 2pk(v2p )2p − 2pkπ −1/p (v2p )2p+2 . dt

(5.152)

We can now show that the set B2p (a) is positively invariant if a ≥ π 1/2p . Proceeding by contradiction, assume that for some g ∈ B2p (a) there is t1 > 0 such that u(t1 , ·) = S(t1 )g ∈ / B2p (a). From (5.152), written as d (v2p )2p ≤ 2pk(v2p )2p 1 − π −1/p (v2p )2 , dt we deduce that v2p is decreasing on any interval where v2p (t) ≥ π 1/2p . Since v2p (t1 ) = |u(t1 , ·)|2p > a , v2p (0) = |u(0, ·)|2p = |g|2p ≤ a , there is an interval [t1 −r,t1 +r] such that v2p (t) ≥ a for |t −t1 | ≤ r, and v2p (t1 −r) = a. But since a ≥ π 1/2p , this implies the contradiction a < v2p (t1 ) ≤ v2p (t1 − r) = a .

5.8

Semilinear Evolution Equations in One Space Dimension

233

Hence, each set B2p (a), with a ≥ π 1/2p , is positively invariant as claimed. We can then conclude. Since π 1/2p a ≥ π 1/2p if a ≥ 1, each set B2p (π 1/2p a) is also invariant. As a consequence of lemma 5.47, the set B∞ (a), being the intersection of positive invariant sets, is positively invariant if a ≥ 1. Finally, we have the following result. PROPOSITION 5.49 If a > 1, B∞ (a) is absorbing with respect to S. PROOF We rewrite (5.152) as 1+1/p d . (v2p )2p ≤ 2pk(v2p )2p − 2pkπ −1/p (v2p )2p dt This is an inequality of the form y0 ≤ αy − β yγ , with α = 2pk, β = 2pkπ −1/p , and γ = 1 + 1p > 1. From (5.153) we deduce y−γ y0 =

1 d 1−γ y ≤ αy1−γ − β , 1 − γ dt

d 1−γ y ≥ α(1 − γ)y1−γ − β (1 − γ) ; dt thus, the function w := y1−γ satisfies the inequality w0 + α(γ − 1)w ≥ β (γ − 1) . Consequently, for t > 0, w(t) ≥ w(0)eα(1−γ)t + αβ 1 − eα(1−γ)t ; from this, recalling that 1 − γ < 0, we deduce that 1/(1−γ) . y(t) ≤ y(0)1−γ eα(1−γ)t + αβ (1 − eα(1−γ)t ) Since 1 − γ = − 1p , α(1 − γ) = −2k, and

β α

= π −1/p , we conclude that

−1/2 −2 v2p (t) ≤ ((v2p (0)) e−2kt + π −1/p 1 − e−2kt ) −1/2 −2 = π −1/p + e−2kt ((v2p (0)) − π −1/p ) .

(5.153)

234

5

Inertial Manifolds

Thus, if u(0, ·) = u0 , −1/2 2 −1/p − π =: b2p (t) . |u(t, ·)|2p ≤ π −1/p + e−2kt |u0 |− 2p

(5.154)

We have to prove that if |u0 |∞ > a, there is Ta > 0 such that |u(t, ·)|∞ ≤ a for all t ≥ Ta . To obtain this, we first show that there is p0 ≥ 1 such that for all p ≥ p0 and t ≥ 0, b2p (t) ≥ π 1/2p .

(5.155)

Indeed, it is immediate to see that, for each t ≥ 0, b2p (t) ≥ π 1/2p if and only if |u0 |2p ≥ π 1/2p . But by (5.150), recalling that a > 1 we deduce that there is p0 ≥ 1 such that for all p ≥ p0 , |u0 |2p ≥ a ≥ π 1/2p ; hence, (5.155) holds. Consequently, by the first part of proposition 5.48, the set B2p (b2p (t)) is positively invariant. Now, (5.150) also implies that −1/2 2 lim b2p (t) = 1 + e−2kt |u0 |− =: b∞ (t) , ∞ −1 p→∞

pointwise in t. As in the second part of the proof of lemma 5.47, we can then show that for all t ≥ 0, \

B2p (b2p (t)) ⊆ B∞ (b∞ (t)) .

(5.156)

p≥1

Consequently, (5.156) and (5.154) imply that for all t ≥ 0 and p ≥ 1, u(t, ·) ∈ B2p (b2p (t)) ⊆ B∞ (b∞ (t)) , that is, |u(t, ·)|∞ ≤ b∞ (t) .

(5.157)

Since b∞ (t) decreases monotonically to 1 as t → +∞, and b∞ (0) = |u0 |∞ > a > 1, there is Ta > 0 such that b∞ (t) ≤ a for all t ≥ Ta . Hence, from (5.157) we conclude that |u(t, ·)|∞ ≤ a if t ≥ Ta . This proves that B∞ (a) is absorbing if a > 1.

5.8.3 Adjusting the Nonlinearity In this section we show how to modify the nonlinearity F(u) = k(u − u3 ) so as to make it globally bounded and Lipschitz continuous from X into X . Before doing so, we show that F is locally Lipschitz continuous, with respect to the L2 norm, from L∞ (0, π) into X . PROPOSITION 5.50 Let a > 0. For all u, v ∈ B∞ (a), kF(u) − F(v)k ≤ k(1 + 3a2 )ku − vk .

5.8

Semilinear Evolution Equations in One Space Dimension

235

PROOF The result is an immediate consequence of the estimate kF(u) − F(v)k2 = k2

Z π

|(u − v) − (u3 − v3 )|2 dx

0

= k2

Z π

|u − v|2 |1 − (u2 + uv + v2 )|2 dx

0

≤ k2 (1 + 3a2 )2

Z π

|u − v|2 dx .

(5.158)

0

To modify F, we fix a > 1 and choose a function b ∈ C∞ b (R), such that |b(r)| ≤ 2a − 1, |b0 (r)| ≤ 1 for all r ∈ R, and b(r) = r

if

|r| ≤ a .

(5.159)

Then, we set Fa (u) := k(b(u) − (b(u))3 ), and claim: PROPOSITION 5.51 Let a > 1. Then Fa is globally bounded and Lipschitz continuous from X into X , with Lipschitz constant `a := k(1 + 3(2a − 1)2 ). PROOF The boundedness of Fa in X = L2 (0, π) follows from the boundedness of b on R. Similarly (omitting the dependence of u and v on x), as in (5.158) Z π

|Fa (u) − Fa (v)|2 dx

0

=k

2

Z π

|b(u) − b(v) − (b(u))3 − (b(v))3 )|2 dx

0

= k2

Z π

|b(u) − b(v)|2 |1 − (b(u))2 + b(u)b(v) + (b(v))2 |2 dx

0

≤k

2

1 + 3(2a − 1)2

2 Z

π

|b(u) − b(v)|2 dx

0

≤ k2 1 + 3(2a − 1)2

2 Z

π

|u − v|2 dx ,

0

having recalled that |b0 (r)| ≤ 1 for all r ∈ R.

5.8.4 The Inertial Manifold We are now in a position to apply theorem 5.42 to the modified equation ut − uxx = Fa (u) .

(5.160)

236

5

Inertial Manifolds

To this end, we recall that the eigenvalues of the operator A := −

d2 dx2

in L2 (0, π), with domain H2 (0, π) ∩ H10 (0, π), are λn = n2 , n ∈ N>0 , and the corresponding eigenfunctions are the functions wn (x) = sin(nx). Thus, the spectral gap condition (5.83) reads (5.161) (N + 1)2 − N 2 = 2N + 1 > 2`a = 2k 1 + 3(2a − 1)2 , and it can obviously be satisfied for sufficiently large N. Moreover, Re λN = N 2 ≥ 0 for all N, so that, as observed in remark 5.43, condition (5.104) also holds. It follows that the semiflow Sa generated by the modified equation (5.160) admits an inertial manifold of the form Ma = graph(ma ) ∩ B∞ (a) , where ma ∈ G`a . Since, by (5.159), Fa (u) = F(u) if |u| ≤ a, it follows that the restriction of the semiflow Sa to Ma coincides with that of the “original” semiflow S on Ma . Hence, Ma is also the desired inertial manifold for S. To see this, let u0 ∈ X . By the smoothing effect of the parabolic operator (see section 3.3.3), we have that S(t)u0 ∈ H1 (0, π) ,→ L∞ (0, π) for all t > 0. Since the set B∞ (a) is absorbing for Sa , there is Ta > 0 such that ua (t, ·) := Sa (t)u0 ∈ B∞ (a) for all t ≥ Ta . Then, since |ua (t, x)| ≤ a, (5.159) implies that ua solves the equation of (5.149). By the uniqueness of solutions of this initial-boundary value problem (with initial values at t = Ta ), it follows that S(t) = Sa (t) on [Ta , +∞[. In particular, for t ≥ Ta , d(S(t)u0 , Ma ) = d(Sa (t)u0 , Ma ) ≤ Ca e−ηt . Finally, we recall that the dimension of Ma is N, as determined by (5.161); thus, N depends on a, via `a . Since inf `a = `1 = 4k , a>1

condition (5.161) yields the lower bound N0 = b4k + 21 c

(5.162)

for the dimension of the inertial manifold, i.e., if N ≥ N0 then theorem 5.42 yields the existence of an N-dimensional inertial manifold for the modified equation (5.160). In fact, it is possible to obtain a better lower bound, in the following way. For r > 0 to be determined, we rewrite the differential equation in (5.149), with f = 0, as ut − uxx + kru = k((1 + r)u − u3 ) . We set Fra (u) := k(b((1 + r)u) − (b(u))3 ) and consider the modified equation ut − uxx + kr u = Far (u) .

(5.163)

5.8

Semilinear Evolution Equations in One Space Dimension

237

Then, as in proposition 5.51, Far is again globally bounded and globally Lipschitz continuous, but now the Lipschitz constant of Far is `ar = k max{1 + r, |3a2 − r − 1|} =: k`(a, r) .

(5.164)

To see this, we proceed as in the proof of proposition 5.51; in particular, it is sufficient to estimate kFar (u) − Far (v)k when both |u|∞ ≤ a, |v|∞ ≤ a. Let h := u2 + uv + v2 . Then, h(x) ≥ 0 for all x ∈ [0, π]. Moreover: if h(x) ≤ 1 + r, |1 + r − h(x)| = 1 + r − h(x) ≤ 1 + r , while if h(x) ≥ 1 + r, |1 + r − h(x)| = h(x) − 1 − r ≤ 3a2 − 1 − r . Consequently, for all x ∈ [0, π] |1 + r − h(x)| ≤ `(a, r) . For u, v ∈ B∞ (a) we compute then that Z π 0

|Far (u) − Far (v)|2 dx = k2

Z π

|u − v|2 |1 + r − h|2 dx ≤ (k`(a, r))2

0

Z π

|u − v|2 dx ,

0

from which (5.164) follows. We turn then to the spectral gap condition for the operator d2 Ar := − 2 + kr dx which appears in (5.163). Since the eigenvalues of Ar are the numbers λ j = j2 + kr, j ≥ 1, the spectral gap condition (5.161) now reads 2N + 1 > k`(a, r) .

(5.165)

As before, we wish to minimize the right side of (5.165). Letting a & 1, we have a first lower bound, given by k`(1, r); in turn, this is minimized by r = 21 , with `(1, 21 ) = 23 . Thus, from (5.165) we deduce the lower bound for the dimension of the inertial manifold N1 := b 32 k + 12 c , i.e., if N ≥ N1 then theorem 5.42 yields the existence of an N-dimensional inertial manifold for the modified equation (5.163). This bound is obviously lower than the bound N0 obtained in (5.162).

238

5

Inertial Manifolds

5.8.5 The Hyperbolic Perturbation In this section we briefly consider the hyperbolic perturbation of problem (5.149), i.e. the one-dimensional IBVP  3   εutt + ut − uxx + k(u − u) = f (x) (5.166) u(0, x) = u0 (x) , ut (0, x) = u1 (x)   u(t, 0) = u(t, π) = 0 . Problem (5.166) is a version of (3.4), with, as in the parabolic case (5.149), n = 1. Thus, we know that (5.166) generates a semiflow S, both in X = H10 (0π) × L2 (0, π) and in X1 = H2 (0, π) ∩ H10 (0, π) × H10 (0, π) . We also know that S admits a bounded, positively invariant absorbing set in both these spaces, as well as a global attractor A ⊆ X , which is contained and bounded in X1 ; S also admits an exponential attractor E ⊆ X . We now proceed to show that theorem 5.44 can be applied to problem (5.166); therefore, S also admits an inertial manifold in X . To this end, as in section 5.7.2 we first rescale the time variable and then transform the equation into a first order system of the form (5.113), with A given by (5.115), and F(U) = (0, f − k(u − u3 )) if U = (u, v) ∈ X . Thus, we need to verify conditions (5.120) and (5.121), after overcoming the difficulty that F is neither globally bounded nor globally Lipschitz continuous on X . To this end, we proceed exactly as in section 5.8.3, by modifying F outside of a bounded, positively invariant absorbing set B (which we know to exist). To do so, we note that the definition of F only involves the first component u of U = (u, v); now, since the projection B˜ of B into H10 (0, π) is bounded, and H10 (0, π) ,→ L∞ (0, π), it follows that B˜ is contained in a ball B∞ (a) of L∞ (0, π). Let a be the radius of this ball and, as in section 5.8.3, set `a := k(1 + 3(2a − 1)2 ), ga (u) := b( f + k(u − u3 )), and Fa (U) := (0, ga (u)), for U = (u, v) ∈ X . Recalling that λN = N 2 , conditions (5.120) and (5.121) read, respectively, 2m + 1 > 4`a , α 2 > (N + 1)2 . The first of these is certainly satisfied for all m such that m ≥ b2`a + 12 c =: N ;

(5.167)

thus, if α is so large that α 2 > (N + 1)2 , the semiflow Sa generated by the modified equation εutt + ut − uxx = ga (u) admits an inertial manifold Ma ⊆ X . Since, by (5.159), Fa (U) = F(U) if U = (u, v) with |u| ≤ a, it follows that the restriction of the semiflow Sa to Ma coincides with

5.8

Semilinear Evolution Equations in One Space Dimension

239

that of the “original” semiflow S on Ma . Hence, Ma is also the desired inertial manifold for S. To see this, let U0 = (u0 , v0 ) ∈ X , and Ua (t) = (ua (t), va (t)) := Sa (t)U0 . Since B is absorbing, there is Ta > 0 such that ua (t, ·) ∈ B∞ (a) for all t ≥ Ta . Then, since |ua (t, x)| ≤ a, (5.159) implies that ua solves the equation of (5.166). By the uniqueness of solutions of this initial-boundary value problem (with initial values at t = Ta ), it follows that S(t) = Sa (t) on [Ta , +∞[. In particular, for t ≥ Ta , d(S(t)U0 , Ma ) = d(Sa (t)U0 , Ma ) ≤ Ca e−ηt . Finally, we recall that the dimension of Ma is N0 , as determined by (5.167). Thus, N0 depends on a, via `a , and we can obtain lower bounds of the dimension by minimizing `a , with a > 1. Since `a is minimized by a = 1, with `1 = 4k, (5.167) implies the lower bound N0 = 8k for the dimension of the inertial manifold.

5.8.6 Concluding Remarks 1. The results of this section 5.8 show that when the perturbation parameter ε is small, the asymptotic properties of the semiflow generated by problems (5.166) are qualitatively similar to those of the semiflow generated by the limit problem when ε = 0, i.e. by (5.149). It would in fact be interesting to further investigate this question, for example in the framework of the upper and lower semicontinuity of these inertial manifolds as ε → 0, as we did in section 3.6 for the attractors. As regards to this, a first difficulty is, of course, that, in general, the inertial manifolds are, in contrast to the global attractors, only positively invariant, and not necessarily invariant.

2. The construction of an inertial manifold for the Chafee-Infante equations with the method described in section 5.6 requires the verification of the spectral gap condition (5.83). In one dimension of space, this is a consequence of (5.161), which guarantees that the eigenvalues of the Laplacian admit arbitrarily large gaps. When n > 1, however, one only knows that, for general domains Ω ⊂ Rn , λN = O(N 2/n ) as N → +∞ (see e.g. Robinson, [Rob93, Rob96]); therefore, it is not known if the point spectrum of the operator contains large gaps. For example, already in the twodimensional case the condition λN = O(N) is clearly not enough to guarantee the existence of large gaps. On the other hand, the situation may be better, if Ω has some particular geometric properties. For instance, the spectral gap condition is satisfied for rectangular domains (see Robinson, [Rob01, sct. 15.4.2]). Similarly, in [MPS88] Mallet-Paret and Sell prove the strong squeezing property for parallelepipeds in R3 , using a technique of spatial averaging. This means that, in this case, one has some type of degenerate spectral gap condition, which can be satisfied without large gaps. Since the method of spatial averaging is a consequence of general properties of the Laplace operator on a domain, the strong squeezing property also holds on some nonrectangular domains in R3 .

240

5

Inertial Manifolds

3. Once again, these difficulties in constructing inertial manifolds may explain the interest in exploring, at least in some cases, the existence of other types of attracting sets which, as in the case of the exponential attractor, retain most of the properties of an inertial manifold; in particular, the finite dimensionality, and the exponential convergence of the orbits. 4. The notion of inertial manifold can be translated and extended to more general classes of differential equations, such as nonautonomous differential equations (see e.g. [GV97, WF97, LL99]), retarded parabolic differential equations (see e.g. [TY94, BdMCR98]), or differential equations with random or stochastic perturbations (see e.g. [Chu95, BF95, CL99, CS01, DLS03]). An extension to abstract nonautonomous dynamical systems, which includes the systems generated by all the equations mentioned above, can be found in [KS02b, KS02a, KS03].

Chapter 6 Examples

In this chapter we present four examples of semiflows, generated by some specific initial-boundary value problems from mathematical physics. These semiflows admit a global attractor in a suitable phase space X ; except for one case, we are also able to show that they also admit an exponential attractor, or even an inertial manifold, in X . The models we consider are: 1. The hyperbolic perturbations of the viscous and nonviscous C AHN -H ILLIARD equations, whose corresponding semiflows admit a global attractor, an exponential attractor and an inertial manifold; 2.

EXTENSIBLE BEAM equation, whose corresponding semiflow admits a global attractor, an exponential attractor and, for special types of forcing terms, an inertial manifold; analogous results are valid for a model of the VON K ÁRMÁN equation in one dimension of space;

3. The NAVIER -S TOKES equations in two dimension of space, whose corresponding semiflow admits a global attractor and an exponential attractor, but the existence of an inertial manifold is open; 4. M AXWELL’s equations in a ferromagnetic medium, for which we can only show the existence of a global attractor. The PDEEs in the first two models are dissipative hyperbolic; the Navier-Stokes equations are essentially parabolic, due to the presence of a viscosity term, while the quasi-stationary Maxwell equations we consider are transformed into a system of quasilinear parabolic equations. All these models can be treated with a suitable application of the methods described in chapters 3, 4 and 5. For other examples of semiflows generated by nonlinear PDEEs, we refer to the books by Temam, [Tem83]; Sell-You, [SY02]; Eden, Foias, Nicolaenko and Temam, [EFNT94], and Constantin, Foias, Nicolaenko and Temam, [CFNT89].

6.1 Cahn-Hilliard Equations In this section we consider the singular perturbations of two IBV problems, concerning respectively the viscous and the nonviscous C AHN -H ILLIARD equations, in

241

242

6

Examples

one dimension of space. Resorting to the techniques presented in the previous chapters, we will show that, at least when the perturbation parameter is sufficiently small, the semiflows generated by these two problems admit global attractors, exponential attractors and inertial manifolds in a suitable phase space. The material we present is mostly taken from Zheng-Milani, [ZM03, ZM05].

6.1.1 Introduction 1. The equations we consider have the unified form εutt + ut + ∆(∆u − u3 + u − δ ut ) = 0 ,

(6.1)

where ε > 0, δ ≥ 0, u = u(t, x), with t > 0 and x ∈ Ω := ]0, π[, and ∆ := ∂ 2 /∂ x2 . We assume that u satisfies the homogeneous Dirichlet boundary conditions u(t, 0) = u(t, π) = 0 , ∆u(t, 0) = ∆u(t, π) = 0 ,

t ≥ 0.

(6.2)

2. When ε = 0, (6.1) yields the classical viscous and nonviscous Cahn-Hilliard equations, corresponding respectively to the cases δ > 0 and δ = 0; these equations are parabolic, and are considered together with the initial condition u(0, x) = u0 (x), x ∈ ]0, π[. This case has been extensively studied; in particular, we refer to Temam, [Tem83, ch. III.4.2], or Sell-You, [SY02, ch. 5.5], and the references cited therein. Other references can be found in Zheng-Milani, [ZM03]. In summary, it is known that the Cahn-Hilliard equations generate a semiflow in the space H := L2 (0, π), and that these semiflows admit a compact global attractor and an inertial manifold in H (for the viscous case, this was proven in Zheng-Milani, [ZM05]). Moreover, the global attractors are lower- and upper-semicontinuous as δ → 0. We are not aware of any result on exponential attractors for either problem; nevertheless, since both semiflows admit a compact absorbing set (because they admit a global attractor), as well as a closed inertial manifold, they also admit an exponential attractor. As we mentioned in the introduction to chapter 5, this exponential attractor is given by the intersection of the absorbing ball with the inertial manifold. 3. When ε > 0 (which is the case we present here), equation (6.1) is hyperbolic; indeed, it is an example of a nonlinear beam equation with viscous damping, of the type we consider in the next section 6.2. Thus, we impose the initial conditions u(0, x) = u0 (x) ,

ut (0, x) = u1 (x)

x ∈ ]0, π[ .

In the sequel, we shall refer to the IBV problem (6.1)+ (6.3)+(6.2) as problem

(CHεδ ) ,

with, for simplicity, ε ∈ ]0, 1] and δ ∈ [0, 1].

(6.3)

6.1

243

Cahn-Hilliard Equations

4. To define the phase spaces in which we study problem (CHεδ ), we introduce the following notations. We set H0 := L2 (0, π), with the usual notations h · , · i and k · k for the scalar product and norm. For integer m ≥ 1, we set 0

−m := (Hm Hm := Hm (0, π) ∩ H10 (0, π) , H0m := Hm 0) , 0 (0, π) , H

and denote by k · km the norm in Hm , m ∈ Z. Because of Poincaré’s inequality, we can choose in H1 the norm kuk1 = k∇uk. Finally, for 1 ≤ p ≤ +∞ we set L p := L p (0, π), and denote by | · | p its norm. The phase spaces we consider are then the following: X0 := H1 × H−1 , X1 := H2 × L2 , X2 := Y × H1 , X−1 := H−1 × Y 0 , (6.4) where Y := {u ∈ H1 : − ∆u ∈ H1 }. Note that H03 ,→ Y ,→ H3 (the second inclusion being a consequence of standard elliptic theory); thus, Y 0 ,→ H−3 . We will consider in X0 and X1 two equivalent norms, which we define in (6.8) and (6.11), (6.12) below. 5. We conclude this introductory part by mentioning some results that we will need in the sequel. At first, we recall (see section A.5.5) that −∆ is an unbounded operator in L2 (0, π), with domain H2 . In particular, −∆ is an isomorphism between Hm and Hm−2 , m ∈ N; we denote its inverse by (−∆)−1 . We consider then the equation formally obtained from (6.1) by taking (−∆)−1 , that is, the equation ε(−∆)−1 utt + (−∆)−1 ut − ∆u + u3 − u + δ ut = 0 .

(6.5)

For u and v ∈ H−1 we set [u, v] := hv, (−∆)−1 uiH−1 ×H1 ;

(6.6)

then, [u, u] = kuk2−1 , and, letting ϕ := (−∆)−1 u and ψ := (−∆)−1 v, [u, v] = h−∆ψ, ϕi = h∇ψ, ∇ϕi ≤ k∇ψk · k∇ϕk = kψk1 · kϕk1 = kuk−1 · kvk−1 . Finally, we recall that, since N = 1, the continuous imbedding H1 (0, π) ,→ L∞ (0, π) holds; we reserve the letter K to denote a constant such that the inequalities kuk−1 ≤ Kkuk ≤ K 2 k∇uk , |u| p ≤ Kkuk1 = Kk∇uk ,

(6.7)

hold for all u ∈ H1 , and 1 ≤ p ≤ +∞. Without loss of generality, we can assume that K ≥ 1. 6. We now introduce in X0 an equivalent norm, whose square is defined by E0 (u, v) := εkvk2−1 + ε[u, v] + 12 kuk2−1 + k∇uk2 ,

(u, v) ∈ X .

(6.8)

244

6

Examples

Since we assume that ε ≤ 1, E0 does define a norm: indeed, by (6.7) we immediately derive that for all (u, v) ∈ X , 2 2 2 2 1 , (6.9) 2 εkvk−1 + k∇uk ≤ E0 (u, v) ≤ α εkvk−1 + k∇uk with α := max{ 32 , K 4 + 1} .

(6.10)

Similarly, we immediately verify that we can consider in X1 the equivalent norms whose squares are defined by E1 (u, v) := εkvk2 + ε hu, vi + 12 kuk2 + k∆uk2 ,

(6.11)

if δ > 0, or, if δ = 0, by 1 E1ε (u, v) := εkvk2 + hu, vi + 2ε kuk2 + k∆uk2 .

(6.12)

6.1.2 The Cahn-Hilliard Semiflows In this section, we show that problem (CHεδ ) generates a semiflow S, both in X0 and in X1 . In some cases, the situation is slightly different in the viscous and the nonviscous cases; for example, when δ = 0 some results are valid only under some limitation on the size of ε. To reduce the length of our presentation, in these cases we will only give a proof for the nonviscous problem, which is more difficult, and refer to Zheng-Milani, [ZM03, ZM05], for the viscous one. 1. At first, we recall the following global existence, uniqueness and regularity result. THEOREM 6.1 For all ε ∈ ]0, 1], δ ∈ [0, 1] and (u0 , u1 ) ∈ X0 , there exists a unique function u ∈ Cb ([0, +∞[; H1 ) ∩ C1b ([0, +∞[; H−1 ), which is a weak solution of problem (CHεδ ) (i.e. with (6.5) satisfied in H−1 , almost everywhere in t). If in addition (u0 , u1 ) ∈ X1 , then u ∈ Cb ([0, +∞[; H2 ) ∩ C1b ([0, +∞[; H0 ) .

(6.13)

PROOF As usual, we limit ourselves to establish formal a priori estimates on weak solutions of problem (CHεδ ). At first, we consider the function Φ0 : X0 → R defined by Φ0 (u, v) := E0 (u, v) + 12 |u|44 − kuk2 + 21 δ kuk2 ,

(6.14)

and note that Φ0 is bounded from below. Indeed, by Minkowski’s inequality, for all η > 0 there is Cη > 0 such that for all u ∈ L4 (0, π), kuk2 ≤ Cη + η|u|44 ;

(6.15)

6.1 hence, taking e.g. η =

1 2

Cahn-Hilliard Equations

245

and recalling (6.9),

Φ0 (u, v) ≥ E0 (u, v) −C1/2 ≥ −C1/2 .

(6.16)

Let now M1 := C2/3 , as defined in (6.15). We claim that for all t ≥ 0, Φ0 (u(t), ut (t)) ≤ (Φ0 (u0 , u1 ) − αM1 ) e−t/α + αM1 ,

(6.17)

where α is as in (6.10). To show this, we begin by multiplying equation (6.5) in H by 2ut and u, and adding the resulting identities. Recalling (6.14), we obtain d Φ0 (u, ut ) + (2 − ε)kut k2−1 + k∇uk2 + |u|44 − kuk2 + 2δ kut k2 = 0 . dt

(6.18)

From this, recalling that ε ≤ 1, we deduce that d Φ0 (u, ut ) + εkut k2−1 + k∇uk2 + |u|44 ≤ kuk2 . dt From (6.14) and (6.9), since also δ ≤ 1, Φ0 (u, ut ) ≤ α εkut k2−1 + k∇uk2 + 21 |u|44 − 12 kuk2 ;

(6.19)

thus, from (6.18) and (6.19), recalling (6.15) and that α ≥ 23 , we obtain d 1 Φ0 (u, ut ) + α1 Φ0 (u, ut ) + 2α kuk2 + 23 |u|44 ≤ kuk2 ≤ C2/3 + 23 |u|44 . dt From this, we conclude that d Φ0 (u, ut ) + α1 Φ0 (u, ut ) ≤ M1 , dt and (6.17) follows by integration. As a consequence of (6.17), we deduce the boundedness of the function t 7→ (u(t), ut (t)) in X0 . Indeed, we immediately deduce that there exists M2 > 0 such that for all t ≥ 0, E0 (u(t), ut (t)) ≤ M2 .

(6.20)

Finally, the proof of the regularity result (6.13) follows from additional a priori estimates, similar to those we establish in proposition 6.4 below. 2. Theorem 6.1 allows us to define the solution operator Sεδ = (Sεδ (t))t ≥0 associated to problem (CHεδ ), with ε ∈ ]0, 1] and δ ∈ [0, 1]. We now show that Sεδ is in fact a semiflow. PROPOSITION 6.2 Let ε, δ ∈ ]0, 1], or, if δ = 0, ε ∈ ]0, 31 ]. Then, for each t > 0, the operator Sεδ is Lipschitz continuous on X0 .

246

6

Examples

PROOF We consider only the case δ = 0. Let z := u − u˜ denote the difference of two solutions of problem (CHε0 ). Then, z solves the equations εztt + zt + ∆ ∆z − (u3 − u˜3 ) + z = 0 , ε(−∆)−1 ztt + (−∆)−1 zt − ∆z + (u3 − u˜3 ) − z = 0 .

(6.21)

As usual, we multiply (6.21) in H by 2zt and z, and add the resulting identities. Setting h := u2 + uu˜ + u˜2 , we obtain d εkzt k2−1 + ε[zt , z] + 12 kzk2−1 + k∇zk2 + hu3 − u˜3 , zi dt + (2 − ε)kzt k2−1 + k∇zk2 + hu3 − u˜3 , zi = hht z, zi + hz, 2zt + zi .

(6.22)

By (6.20), we can start the estimate of the first term of the right side of (6.22) by 2huut z, zi ≤ kut k−1 kuz2 k1 ≤ C k∇uk|z|2∞ + 2|u|∞ |z|∞ k∇zk , (6.23) where C depends only on M2 of (6.20). Resorting then to the Gagliardo-Nirenberg inequality |z|∞ ≤ Ck∇zk1/2 kzk1/2 +Ckzk (see theorem A.70), we obtain from (6.23) 2 huut z, zi ≤ C k∇zk kzk + k∇zk3/2 kzk1/2 + kzk2 ≤

2 2 1 24 k∇zk +C2 kzk .

The other three terms of hht z, zi can be treated in the same way, leading to the estimate hht z, zi ≤ 16 k∇zk2 +Ckzk2 .

(6.24)

hz, 2zt + zi ≤ 85 kzt k2−1 + 58 k∇zk2 + kzk2 .

(6.25)

Next, we estimate

Calling Ψ (z, zt ) the differentiated term of the left side of (6.22), i.e. Ψ (z, zt ) := εkzt k2−1 + ε[zt , z] + 12 kzk2−1 + k∇zk2 + hu3 − u˜3 , zi ,

(6.26)

by (6.24) and (6.25) we obtain from (6.22) that d 5 Ψ (z, zt ) + ( 52 − ε)kzt k2−1 + 24 k∇zk2 + hu3 − u˜3 , zi ≤ Ckzk2 . dt

(6.27)

Assume now e.g. that ε ≤ 13 . Then, 52 − ε > 16 ε; and since the function u 7→ u3 is monotone, we deduce from (6.27) that d Ψ (z, zt ) + 16 kzt k2−1 + k∇zk2 + hu3 − u˜3 , zi ≤ Ckzk2 . dt

(6.28)

6.1

247

Cahn-Hilliard Equations

It is now easy to verify that, since ε ≤ 1, Ψ (z, zt ) ≤ α εkzt k2−1 + k∇zk2 + hu3 − u˜3 , zi , with α as in (6.10); consequently, we obtain from (6.28) that d 1 Ψ (z, zt ) + 6α Ψ (z, zt ) ≤ Ckzk2 . dt

(6.29)

Integrating (6.29), we obtain that for all t ≥ 0 Ψ (z(t), zt (t)) ≤ Ψ (z(0), zt (0))e−t/6α +C

Z t

kzk2 ds .

(6.30)

0

From (6.26) and (6.9), we deduce that Ψ (z, zt ) ≥

1 2

εkzt k2−1 + k∇zk2 ≥

1 2α E0 (z, zt ) ;

(6.31)

moreover, we also have that

˜ ∞ + |u| ˜ 2∞ kzk2 ≤ C3 k∇zk2 , 0 ≤ u3 − u˜3 , z = hhz, zi ≤ |u|2∞ + |u|∞ |u| where C3 depends only on M2 and K. Hence, recalling that the sum of the first three terms of Ψ is positive definite, and that ε ≤ 1, we also have Ψ (z, zt ) ≤ εkzt k2−1 + ε[zt , z] + 21 kzk2−1 + (1 +C3 + K)k∇zk2 ≤ C4 E0 (z, zt ) . (6.32) Finally, since kuk2 ≤ 4E0 (u, v) for all (u, v) ∈ X0 , we deduce from (6.30), (6.31) and (6.32) that z satisfies the estimate E0 (z(t), zt (t)) ≤ MαE0 (z(0), zt (0))e−t/6α + Mα

Z t 0

E0 (z, zt ) ds ,

(6.33)

with M := max{2C4 , 8C}. Applying Gronwall’s inequality, we deduce from (6.33) that, for all t ≥ 0, E0 (z(t), zt (t)) ≤ M αE0 (z(0), zt (0))eMαt .

(6.34)

Thus, each operator Sε0 (t) is Lipschitz continuous in X0 , as claimed. The proof in the case δ > 0 is similar (and actually simpler).

6.1.3 Absorbing Sets 1. The existence of a bounded, positively invariant absorbing set for the semiflow Sεδ in X0 is an immediate consequence of estimate (6.17). PROPOSITION 6.3 Let ε ∈ ]0, 1] and δ ∈ [0, 1]. For any R0 > αM1 +C1 , the ball B0 := {(u, v) ∈ X0 : E0 (u, v) ≤ R0 }

248

6

Examples

is absorbing for Sεδ . Moreover, for any R > αM1 , the set B0 := {(u, v) ∈ X : Φ0 (u, v) ≤ R}

(6.35)

is bounded, positively invariant and absorbing for Sεδ in X0 . PROOF The first claim follows from (6.16) and (6.17). In particular, for all t ≥ 0, E0 (u(t), ut (t)) ≤ (Φ0 (u0 , u1 ) − αM1 )e−t/α + αM1 +C1 .

(6.36)

Assume now that (u0 , u1 ) is in a bounded set G ⊆ X0 . There exists then Γ ≥ 1 such that E0 (u0 , u1 ) ≤ Γ . Now, from (6.14) and (6.7), recalling also (6.10), Φ0 (u0 , u1 ) ≤ E0 (u0 , u1 ) + 41 |u0 |44 ≤ Γ + 41 K 4Γ 2 ≤ αΓ 2 ; thus, from (6.36) we deduce that for all t ≥ 0, E0 (u(t), ut (t)) ≤ α(Γ 2 − M1 )e−t/α + αM1 +C1 . From this it follows that if α(Γ 2 − M1 ) ≤ R0 − (αM1 +C1 ), then E0 (u(t), ut (t)) ≤ R0 for all t ≥ 0, while if α(Γ 2 − M1 ) > R0 − (αM1 +C1 ), then E0 (u(t), ut (t)) ≤ R0 for all t ≥ TG , with α(Γ 2 − M1 ) TG := α ln . R0 − (αM1 +C1 ) This proves that the ball B0 is absorbing. The boundedness of the set B0 follows from (6.16) and (6.17), and its positive invariance is a direct consequence of (6.17). In fact, if Φ0 (u0 , u1 ) ≤ R, then for all t ≥ 0 Φ0 (u(t), ut (t)) ≤ (R − αM1 )e−t/α + αM1 ≤ (R − αM1 ) + αM1 ≤ R . Finally, we prove that B0 is absorbing exactly in the same way as we did for B0 ; we find that Φ0 (u(t), ut (t)) ≤ R for all t ≥ T˜G , where now  0 if α(Γ 2 − M1 ) ≤ R − αM1 ; 2 ˜ TG := α(Γ − M1 )  if α(Γ 2 − M1 ) > R − αM1 . R − αM1 This concludes the proof of proposition 6.3; we remark that the set B0 is not a ball of X0 . 2. We now show that the semiflow Sεδ also admits an absorbing set in X1 . PROPOSITION 6.4 For each δ ∈ ]0, 1], there exists R1δ > 0, depending on δ but not on ε, such that for all ε ∈ ]0, 1], the set B1δ := {u, v) ∈ X1 ∩ B : E1 (u, v) ≤ R1δ } ,

6.1

249

Cahn-Hilliard Equations

where E1 is defined in (6.11), is positively invariant, bounded and absorbing in X1 for the semiflow Sεδ . Analogously, if δ = 0, there exist ε0 ∈ ]0, 13 ], with the property that for all ε ∈ ]0, ε0 ], there is R1ε such that the set B1ε := {(u, v) ∈ X1 ∩ B : E1ε (u, v) ≤ R1ε } ,

(6.37)

where E1ε is defined in (6.12), is positively invariant, bounded and absorbing in X1 for the semiflow Sε0 . PROOF We only consider the case δ = 0. We multiply equation (6.1) by 2ut and 1 ε u in H. We obtain d E1ε (u, ut ) + kut k2 + ε1 k∆uk2 = − ε1 h∇(u3 − u), ∇ui + 2h∆(u3 − u), ut i. dt Denoting by Cb a generic positive constant, depending on the uniform bound on u in H1 (0, π) ,→ L∞ (0, π), we can estimate − ε1 h∇(u3 − u), ∇ui = − ε1 h(3u2 − 1)∇u, ∇ui ≤ ε1 k∇uk2 ≤ ε1 Cb .

(6.38)

Since ∆(u3 − u) = div((3u2 − 1)∇u) = 6u∇u · ∇u + (3u2 − 1)∆u , resorting to the Gagliardo-Nirenberg inequality |∇u|4 ≤ Ck∆2 uk1/4 k∇uk3/4 +Ck∇uk , and to the elliptic estimate kuk2 ≤ Ck∆uk +Ckuk (see theorems A.70 and A.77), we have ku ∇u · ∇uk ≤ |u|∞ |∇u|24 ≤ Cb (1 + kuk2 ) ≤ Cb (1 + k∆uk) , k(3u2 − 1)∆uk ≤ (3|u|2∞ + 1)k∆uk ≤ Cb k∆uk . Consequently (for different Cb ),

2 ∆(u3 − u), ut ≤ 12 kut k2 +Cb +Cb k∆uk2 .

(6.39)

Choose now ε0 so small that 2ε0Cb ≤ 1. Then, if ε ≤ ε0 , by (6.39) and (6.38) we obtain from (6.38) d 1 E1ε (u, ut ) + 2ε εkut k2 + k∆uk2 ≤ 1 + ε1 Cb . dt Since, as we easily verify, E1ε (u, ut ) ≤

3 2

εkut k2 + k∆uk2 + ε1 kuk2 ,

(6.40)

250

6

Examples

we deduce from (6.40) that, if also ε ≤ 21 , d 1 E1ε (u, ut ) ≤ ε2 Cb + ε12 kuk2 ≤ E1ε (u, ut ) + 3ε dt

2 C . ε2 b

(6.41)

Let Cε := ε6 Cb . Integrating (6.41), we deduce that for all t ≥ 0, E1ε (u(t), ut (t)) ≤ (E1ε (u0 , u1 ) −Cε ) e−t/3ε +Cε . Thus, it follows that if R1ε > Cε , the set B2ε defined in (6.37) is positively invariant and absorbing for Sε0 . This concludes the proof of proposition 6.4 (when δ = 0); note that, in general, R1ε is unbounded as ε → 0.

6.1.4 The Global Attractor In this section we resort to theorem 2.56 of chapter 2 to show that the semiflow S generated by problem (CHεδ ) admits a global attractor Aεδ in X0 . More precisely, we resort to the α-contraction method described in section 3.4.5, to show that the ω-limit set ωεδ (B0 ), where B0 is the absorbing set determined in proposition 6.3, is the global attractor for Sεδ in X0 . Note that this set is not empty, since it contains the stationary solutions of problem (CHεδ ). Recalling proposition 2.59, to apply theorem 2.56 it is sufficient to find t∗ > 0 such that the operator Sεδ (t∗ ) is an α-contraction in X0 , up to a precompact pseudometric. THEOREM 6.5 Let ε, δ ∈ ]0, 1], or, if δ = 0, ε ∈ ]0, 13 ]. The set Aεδ := ωεδ (B0 ) is a global attractor in X0 for the semiflow Sεδ generated by problem (CHεδ ). PROOF It is sufficient to note that, as a second consequence of estimate (6.33), we can apply theorem 2.56 to the semiflow Sεδ . Indeed, if e.g. we choose t∗ > 0 such that q∗ := M αe−t∗ /6α < 1, the operator Sεδ (t∗ ) is a strict contraction in X0 , up to the pseudometric ψ∗ defined by Z ψ∗ ((u, v), (u, ˜ v)) ˜ := αM

t∗

1/2 kz(s)k2 ds ,

0

where, for (u, v), (u, ˜ v) ˜ ∈ X0 , z := u − u˜ is the difference of the solutions to problem (CHεδ ), corresponding to initial values (u, v) and (u, ˜ v). ˜ This pseudometric is clearly precompact, because of the compactness of the injection {u ∈ L2 (0,t∗ ; H1 ) : ut ∈ L2 (0,t∗ ; H−1 )} ,→ L2 (0,t∗ ; L2 (0, π)) . Thus, by proposition 2.59, the map Sεδ (t∗ ) is an α-contraction. In turn, theorem 2.56 implies that the ω-limit set of the set B0 defined in (6.35) is the global attractor for the semiflow Sεδ .

6.1

Cahn-Hilliard Equations

251

REMARK 6.6 We can say a lot more on the dependence of the attractors Aεδ on the parameters ε and δ . In particular, we can show that Aεδ is uniformly bounded, in X0 , with respect to both parameters ε and δ . Moreover, if δ > 0, the attractor Aεδ is contained in a bounded set of X2 , which is independent of ε. The additional regularity of the attractor when δ > 0 is due to the presence of the term −δ ∆ut , which has a regularizing effect on the solution. In contrast, the question of the validity of the analogous result for Aε0 is open; in fact, we do not even know if the inclusion Aε0 ⊂ X1 holds. Finally, the attractors Aεδ are also upper-semicontinuous with respect to both ε and δ . More precisely, let A0δ ⊂ H3 be the global attractor of the semiflow S0δ generated by the parabolic problem (CH0δ ), and define the set A0δ := {(u, v) ∈ X0 : u ∈ A0δ , v = −∆(I − δ ∆)−1 (u − u3 + ∆u)} . We interpret A0δ as the “natural” imbedding of A0δ in X0 . Then, we have the commutative diagram Aεδ ↓ A0δ

−→ −→

Aε0 ↓ , A00

where the vertical arrows mean convergence as ε → 0, and the horizontal arrows mean convergence as δ → 0, in the sense of the semidistance ∂ of (2.2), in the topology of X1 . For a proof of these results, based on techniques analogous of those we used in sections 3.5 and 3.6, we refer to Zheng-Milani, [ZM03].

6.1.5 The Exponential Attractor In this section we show that the semiflow Sεδ also admits an exponential attractor Eεδ in X0 , which contains the global attractor Aεδ . THEOREM 6.7 In the same conditions of proposition 6.4, the semiflow Sεδ generated by problem (CHεδ ) admits an exponential attractor Eεδ in X0 . PROOF As before, we prove theorem 6.7 in detail only in the case δ = 0. We apply theorem 4.5 of chapter 4. To this end, we consider the absorbing sets B1δ or B1ε for Sεδ in X1 , determined in proposition 6.4. Since the injection X1 ,→ X0 is compact, these sets are compact in X0 . We propose then to show that Sεδ satisfies the discrete squeezing property (see definition 4.3), relative to the set B1 , where B1 = B1δ if δ > 0, and B1 = B1ε if δ = 0. 1. We proceed almost exactly as in the proof of theorem 4.10 of section 4.4.2, of which we keep the same choices of N and XN , and denote by k · k0 the norm induced on X0 by E0 . Thus, we show that, given any t∗ > 0 and γ ∈ ]0, 12 [, there exists an

252

6

Examples

integer N∗ , with the property that if u0 , u¯0 ∈ B1 are such that S(t∗ )u0 − S(t∗ )u¯0 ∈ / CN∗ , i.e. if kPN∗ (S(t∗ )u0 − S(t∗ )u¯0 )k0 < kQN∗ (S(t∗ )u0 − S(t∗ )u¯0 )k0

(6.42)

(that is, if (4.13) holds for the operator S(t∗ )), then (4.11) must hold, i.e. kS(t∗ )u0 − S(t∗ )u¯0 k0 ≤ γku0 − v0 k0 .

(6.43)

To this end, we define on X0 the function 2

2

δ M0 (u, v) := ε kvk−1 + [u, v] + k∇uk + 2ε kuk2 .

It is then easy to verify that, if ελN+1 ≥ K 2 + 1, with K as in (6.7), M0 is the square of an equivalent norm in the subspace QN (X ) (the projection QN being defined as in (4.38)), with 1 2α E0 (u, v) ≤

M0 (u, v) ≤ 3E0 (u, v) .

(6.44)

2. We now estimate the difference of solutions of (6.1), whose orbits are in B1 . If u and u˜ are two such solutions, corresponding to initial values U0 := (u0 , u1 ), U˜ 0 := (u˜0 , u˜1 ) ∈ B1 , we set z(t) := u(·,t) − u(·,t), ˜ and Z(t) := (z(t), zt (t)). At first, we recall estimate (6.34), which provides a control of the growth of Z on bounded time intervals. We can rewrite (6.34) as E0 (z(t), zt (t)) ≤ C E0 (z(0), zt (0))ect , with C and c independent of u, u˜ and t. Next, we establish a linear differential inequality on Z in QN X0 , for N ∈ N so large that ελN+1 ≥ max{K 2 + 1, 4}. More precisely, we set q := QN (z), and claim that the function t 7→ M0 (q(t), qt (t)) satisfies the linear differential inequality d 2 1 M0 (q, qt ) + 3ε M0 (q, qt ) ≤ Kδ k∇z(t)k , dt

(6.45)

where, when δ = 0, the constant K0 is independent of t, N and ε, and, if δ > 0, Kδ := δ λK0 . As stated above, we show (6.45) only for the case δ = 0. Applying QN N+1 to equation (6.21), and noting that QN commutes with −∆, we see that q satisfies the equation ε(−∆)−1 qtt + (−∆)−1 qt − ∆q = q − QN (u3 − u˜3 ) =: gN .

(6.46)

Multiplying (6.46) in L2 (0, π) by 2qt and ε1 q, adding the resulting identities, we obtain d 2 M0 (q, qt ) + kqt k2−1 + ε1 [qt , q] + ε1 k∇qk = ε1 hgN , 2εqt + qi . dt

(6.47)

6.1 Splitting the term

1 ε

Cahn-Hilliard Equations

253

[qt , q] in two, we deduce from (6.47)

d 2 1 1 M0 (q, qt ) + kqt k2−1 + 2ε [qt , q] + ε1 k∇qk ≤ ε1 hgN , 2εqt + qi − 2ε [qt , q] . (6.48) dt Since ε ≤ 1, we have that K2 + 1 2 2 1 k∇qk . k∇qk ≤ 14 kqt k2−1 + 4ε 4ε 2 λN+1

(6.49)

+ qi ≤ C ε2 k∇q − ∇(u3 − u˜3 )k (εkqt k−1 + kqk) .

(6.50)

1 [qt , q] ≤ 14 kqt k2−1 + − 2ε

Next, we estimate 1 ε hgN , 2εqt

Writing ∇(u3 − u˜3 ) = 3u2 (∇u − ∇u) ˜ + 3(u2 − u˜2 )∇u˜ , we have k∇q − ∇(u3 − u˜3 )k ≤ |3u2 − 1|∞ k∇zk + 3|u + u| ˜ ∞ k∇uk|z| ˜ ∞ ≤ Ck∇zk , (6.51) where C depends on the uniform bounds on u and u˜ in H1 . Inserting (6.51) into (6.50), and recalling that ελN+1 ≥ 4, we obtain (for different constants C) ! 1 1 p k∇qk ε hgN , 2εqt + qi ≤ Ck∇zk kqt k−1 + ε λN+1 1 k∇qk2 (6.52) ≤ Ck∇zk2 + 41 kqt k2−1 + 2 ε λN+1 1 ≤ Ck∇zk2 + 14 kqt k2−1 + 4ε k∇qk2 .

From (6.52), (6.49) and (6.48) we deduce then that, for suitable C > 0, d M0 (q, qt ) + 21 kqt k2−1 + ε1 [qt , q] + ε1 k∇qk2 ≤ Ck∇zk2 , dt which yields (6.45) when δ = 0. 3. We now recall (6.34), which provides an estimate on the difference of solutions on bounded intervals. Setting β := Mα, (6.34) yields 2 1 2 k∇z(t)k

≤ E0 (z(t), zt (t)) ≤ β E0 (z(0), zt (0))eβt ,

for all t ≥ 0. Replacing this into (6.45), then integrating, and recalling (6.44), we obtain M0 (q(t), qt (t)) ≤ 3E0 (z(0), zt (0))e−t/3ε + 6β εM0 E0 (z(0), zt (0))eβt

(6.53)

254

6

Examples

(for δ = 0). Given γ ∈ ]0, 12 [, we first choose t∗ > 0 so large that 8α e−t∗ /3ε0 ≤ 1, and then ε1 ∈ ]0, ε0 ] so small that 16αβ ε1C5 eβt∗ ≤ γ 2 . With these choices, we deduce from (6.53), that, if ε ≤ ε1 , M0 (q(t∗ ), qt (t∗ )) ≤

1 2 4α γ E0 (z(0), zt (0)) .

Thus, if (6.42) holds, then, by (6.44), E0 (z(t∗ ), zt (t∗ )) ≤ 2E0 (QN∗ (z(t∗ ), zt (t∗ ))) ≤ 4αM0 (q(t∗ ), qt (∗ )) ≤ γ 2 E0 (z(0), zt (0)) . This means that (6.43) holds, as desired. Consequently, we conclude that if δ = 0 and ε ≤ ε1 , the semiflow Sε0 satisfies the discrete squeezing property. We can then apply theorem 4.5, and conclude the proof of theorem 6.7 when δ = 0. The proof when δ > 0 is similar (and actually simpler).

6.1.6 The Inertial Manifold In this section we show that the semiflow Sεδ also admits an inertial manifold in X0 , constructed with the techniques of section 5.7.2 of chapter 5. That is, the inertial manifold is the graph of a Lipschitz continuous function defined over a finite dimensional subspace of X0 . √ 1. As in section 5.7.2, we introduce the time rescaling t 7→ εt, which transforms equation (6.1) into utt + 2αut + ∆ ∆u − u3 + u − 2αδ ut = 0 , (6.54) with α :=

1 √ . 2 ε

Then, we transform equation (6.54) into the first order system Ut + AU = F(U)

(6.55)

for U = (u, v) ∈ X0 , where A=

0 ∆2

−1 2α(1 − δ ∆)

! , F(U) :=

0 g(u)

! ,

(6.56)

with g(u) := ∆(u3 − u). Note that, if δ > 0, dom(A) = {u ∈ H3 (0, π) : u ∈ H, uxx ∈ H2 } × H1 , while if δ = 0, the second factor is to be replaced by H−1 . System (6.55) generates a semiflow in X , which we still denote by Sεδ : our goal is to show that the semiflow generated by system (6.55) satisfies the strong squeezing property. In turn, this will be a consequence of the fact that the operator A in (6.55) satisfies the spectral gap condition (5.83) of definition 5.32, either with respect to the standard graph norm in X0 , defined in (6.57) below, or to an equivalent one.

6.1

255

Cahn-Hilliard Equations

2. We consider in X0 the usual graph norm, induced by the scalar product hU,V i0 := h∇u, ∇yi ¯ + [¯z, v] , U = (u, v) ,

V = (y, z) ∈ X0 ,

(6.57)

where the bar denotes complex conjugation. Note that the last term of (6.57), defined in (6.6), makes sense, because z ∈ H−1 , and (−∆)−1 v ∈ H1 since v ∈ H−1 . Moreover, the operator A defined in (6.56) is monotone. Indeed, for U ∈ dom(A), hAU,Ui0 is real and nonnegative, since

hAU,Ui0 = −h∇v, ∇ui ¯ + v, ¯ (−∆)−1 (∆2 u + 2α(1 − δ ∆)v) = −h∇v, ∇ui ¯ + h∇v, ¯ ∇ui + 2αkvk2−1 + 2αδ kvk2 = 2α kvk2−1 + δ kvk2 . To determine the eigenvalues of A, we observe that the eigenvalue equation AU = µU ,

U = (u, v) ∈ X0 ,

is equivalent to the system −v = µu ,

∆2 u + 2α(1 − δ ∆)v = µv .

Thus, u must solve the eigenvalue problem  2 2   ∆ u + 2αδ µ∆u = (2α µ − µ )u , u(0) = u(π) = 0 ,   ∆u(0) = ∆u(π) = 0 .

(6.58)

(6.59)

We easily see that (6.59) has, for each positive integer j, the pair of eigenvalues q 2 µ± := α(1 + δ j ) ± α 2 (1 + δ j2 )2 − j4 ∈ C ; j thus, A does have a countable set of eigenvalues, with Re µ ± j > 0 for all j. Because of the first of (6.58), the corresponding eigenfunctions have the form U j± = q 2 (u j , −µ ± j u j ), with u j (x) = π sin( jx). For future reference, we note that for all j≥1 k∇u j k = j ,

ku j k−1 = j−1 .

(6.60)

3. A remarkable feature of system (6.55), with A and F defined in (6.56), is a difference in the distribution of the eigenvalues µ ± j , and the consequent possibility of satisfying the spectral gap condition, according to whether δ > 0 or δ = 0. Indeed, if δ = 0 the eigenvalues reduce to p µ± α 2 − j4 ; j =α±

256

6

Examples

thus, if α < 1, i.e. if ε is so large that 4ε > 1, all the eigenvalues of A are complex, nonreal, and have the same real part α. This situation is similar to the semilinear dissipative wave equation considered in section 5.7.2; as in that example, in this case it is impossible to find any decomposition of the eigenvalues of A such that the spectral gap condition (5.83) holds, and the existence of an inertial manifold for equation (6.55) cannot be guaranteed. On the other hand, when α > 1 some eigenvalues of A will be real, positive and distinct, so that the spectral gap condition may hold. In contrast, we will see that the spectral gap condition always holds if δ > 0, even if all eigenvalues µ ± j are complex, nonreal. 4. In the remainder of this section, we will assume that the nonlinearity g : H1 → H−1 in (6.56) is globally bounded and globally Lipschitz continuous, with Lipschitz constant `. To construct the inertial manifold for the original semiflow Sεδ , we would then need to adjust the nonlinearity, in a way similar to section 5.8.3; for this technical part, we refer to Zheng-Milani, [ZM05]. Under this assumption, in the viscous case δ > 0 we have THEOREM 6.8 Let N1 ∈ N be so large that if N ≥ N1 , then: 1. If 0 < αδ < 1, αδ (2N + 1) > 2`; √ 2. If αδ = 1, 2N ≥ 2` + 2α and p

2αN 2 + α 2 −

q

√ 2α(N + 1)2 + α 2 + 2α ≥ −1 .

3. If αδ > 1, the inequalities p 2` +1, (2N + 1) αδ − α 2 δ 2 − 1 ≥ √ αδ − 1 p p p R(N) − R(N + 1) + (2N + 1) α 2 δ 2 − 1 ≤ 1

(6.61) (6.62)

hold, where R(N) := (α 2 δ 2 − 1)N 4 + 2α 2 δ N 2 + α 2 .

(6.63)

In each of these cases, the operator A satisfies the spectral gap condition (5.83) in X0 , with respect to either the graph norm (6.57), or an equivalent one. Consequently, the semiflow Sεδ generated by (6.55) admits an inertial manifold in X0 , of the form (5.16). Similarly, in the nonviscous case δ = 0 we have

6.1

257

Cahn-Hilliard Equations

THEOREM 6.9 Let N ∈ N be the smallest integer such that (N + 1)4 − N 4 > 4`. Assume that α = is so large that 4`2 α 2 > (N + 1)4 + . (N + 1)4 − 4 `

1 √ 2 ε

There is then an equivalent norm in X such that the operator A defined in (6.56), with δ = 0, satisfies the spectral gap condition. Consequently, the semiflow S generated by (6.55) admits an inertial manifold in X0 , of the form (5.16). 5. We will only prove theorem 6.8 in the case αδ > 1, which is the most difficult; for all other cases, we refer to Zheng-Milani, [ZM05]. Since αδ > 1, all eigenvalues µ ± j of A are real and positive, and we easily see that + ) both sequences (µ − ) and (µ j ≥ 1 j j≥1 are increasing. We will proceed in four steps. j 5.1.

√ Setting γ := αδ + α 2 δ 2 − 1, we easily check that, as j → +∞ 2 µ± j = γ j +α ± √

α 2δ α 2δ 2 − 1

+ O( j−2 ) .

(6.64)

Since γ > 1, (6.64) implies that it is impossible to decompose the point spectrum of A in such a way that the corresponding subspaces X1 and X2 are orthogonal. Indeed, for any such decomposition σ1 ∪ σ2 there is at least one index j such that U j− ∈ X1 + 4 and U j+ ∈ X2 ; for this j, recalling (6.57) and (6.60), and noting that µ − j µ j = j , we compute that D E − + 2 2 2 + U j− ,U j+ = k∇u j k2 + [−µ − j u j , −µ j u j ] = k∇u j k + µ j µ j ku j k−1 = 2 j . X

To overcome this difficulty, we will define a new scalar product in X0 , equivalent to (6.57), with respect to which the subspaces X1 and X2 will be orthogonal. To this end, we note that the same asymptotic distribution (6.64) assures that, if the eigenvalues µ ± j are listed in nondecreasing order, then for arbitrarily large N there − are consecutive eigenvalues µN− and µN+1 . More precisely, PROPOSITION 6.10 Let the eigenvalues µ ± j , j ≥ 1, be arranged in nondecreasing order. For all m ∈ N, − there is N ≥ m such that µN− and µN+1 are consecutive. PROOF For m ≥ 1, let qm denote the number of indices n such that µm+ < µn− ≤ + µm+1 . We have to show that for all m ∈ N there is N ≥ m such that qN ≥ 2. Assume otherwise, i.e., that there is m0 ≥ 1 such that for all m ≥ m0 , qm ≤ 1. This means that + ]. In turn, this defines a function for each m ≥ m0 there is at most one µn− ∈ ]µm+ , µm+1

258

6

Examples

m 7→ n(m), m ≥ m0 . For r ≥ 0, let mr := m0 + r and nr := n(mr ). Then, after possibly a finite number of them, the eigenvalues are ordered as µm+0 < qm0 µn−0 + (1 − qm0 )µm+1 ≤ µm+1 < qm1 µn−1 + (1 − qm1 )µm+2 ≤ · · · ≤ µm+r < qmr µn−r + (1 − qmr )µm+r+1 ≤ µm+r+1 < · · · .

(6.65)

Since the sequence (µ − j ) j≥1 is increasing, (6.65) implies that nr = n0 + r, and, therefore, µn−0 +r > µm+0 +r for all r ≥ 0. Now, from (6.64) we have that 2 1 γ (n0 + r)

> γ(m0 + r)2 + √

2α 2 δ α 2δ 2 − 1

+ O r−2

as r → +∞. However, this is impossible, since γ > 1. − are consecutive, we separate the eigen5.2. Given then N such that µN− and µN+1 values of A as follows. Denoting by σp (A) the point spectrum of A, i.e. the sequence of its eigenvalues (µ ± j ) j∈N>0 , and by U the set of the corresponding sequence of ± eigenvectors (U j ) j∈N>0 , we set

− + I0 = { j ∈ N : µ − j ≤ µ j ≤ µN } ,

− + I1 = { j ∈ N : µ − j ≤ µN < µ j }

− σ1 := {µ ± j : j ∈ I0 } ∪ {µ j : j ∈ I1 } , σ2 := σp (A) \ σ1 ,

U1 = {U j± : j ∈ I0 } ∪ {U j− : j ∈ I1 } ,

U2 = U \ U1 ,

(6.66)

and consider the corresponding decomposition of X1 := span U1 , X2 := span U2 .

(6.67)

We explicitly note that, in this section, X1 and X2 denote the subspaces of X0 defined in (6.67), and not those defined in (6.4) and (6.7). Our goal is to make these two subspaces orthogonal, and to show that the spectral inequality (5.83) holds, with − Λ1 = µN− and Λ2 = µN+1 , in accord with (6.66). We further decompose X2 := XC ⊕ XR , with XC := span UC ,

UC = {U j+ : j ∈ I1 } ,

XR := span UR , UR = U2 \ UC ,

(6.68)

and set XN := X1 ⊕ XC . Note that X1 and XC are finite dimensional, that UN− ∈ X1 , − UN+1 ∈ XR , and that the reason why X1 is not orthogonal to X2 is that, while it is orthogonal to XR , X1 is not orthogonal to XC . We now introduce two functions Φ : XN → R and Ψ : XR → R, defined by D E Φ(U,V ) := 2αhu, yi ¯ + (2αδ − 1)h∇u, ∇yi ¯ + (−∆)−1/2 z¯, (−∆)1/2 u D E + (−∆)−1/2 v, ¯ (−∆)1/2 y + [¯z, v] , (6.69)

259

6.1

Cahn-Hilliard Equations D E Ψ (U,V ) := αδ h∇u, ∇yi ¯ + (−∆)−1/2 z¯, (−∆)1/2 u D E + (−∆)−1/2 v, ¯ (−∆)1/2 y + [¯z, v] ,

(6.70)

with U = (u, v), V = (y, z) ∈ XN or, respectively, XR . These functions are well defined: Indeed, since u ∈ H1 (0, π) and z ∈ H−1 (0, π), then both (−∆)−1/2 z and (−∆)1/2 u are in L2 (0, π), and analogously for y and v. We now show that Φ and Ψ are positive definite. Let first U = (u, v) ∈ XN : then, D E 2 Φ(U,U) = 2αkuk2 + (2αδ − 1)k∇uk2 + 2 (−∆)−1/2 v, ¯ (−∆)1/2 u + kvk−1 2

≥ 2αkuk2 + (2αδ − 1)k∇uk2 − 2kvk−1 k∇uk + kvk−1 2

≥ 2αkuk2 + (2αδ − 1)k∇uk2 − kvk2−1 − k∇uk2 + kvk−1 ≥ 2αkuk2 + 2(αδ − 1)k∇uk2 .

(6.71)

Since αδ > 1, we conclude that Φ(U,U) ≥ 0 for all U ∈ XN . Analogously, for U ∈ XR : D E Ψ (U,U) = αδ k∇uk2 + 2 (−∆)−1/2 v, ¯ (−∆)1/2 u + kvk2−1 2

≥ (αδ − 1)k∇uk2 + k∇uk − 2kvk−1 k∇uk + kvk2−1 2

≥ (αδ − 1)k∇uk ,

(6.72) (6.73)

from which we conclude that also Ψ (U,U) ≥ 0 for all U ∈ XR . Thus, Φ and Ψ define a scalar product, respectively on XN and XR , and we can define an equivalent scalar product in X0 , by hhU,V ii := Φ(PN U, PN V ) +Ψ (PRU, PRV ) ,

(6.74)

where PN and PR are, respectively, the projections of X onto XN and XR . For simplicity, with a slight abuse of notation we shall write (6.74) simply as hhU,V ii := Φ(U,V ) +Ψ (U,V ) .

(6.75)

We proceed then to show that the subspaces X1 and X2 defined in (6.67) are orthogonal with respect to the scalar product (6.75). In fact, it is sufficient to show that X1 is orthogonal to XC ; in turn, this reduces to showing that hhU j− ,U j+ iiX = 0 if U j− ∈ XN and U j+ ∈ XC . Recalling (6.69) and (6.70), we immediately compute that hhU j− ,U j+ ii = Φ(U j− ,U j+ ) − 2 = 2αku j k2 + (αδ − 1)k∇u j k2 − (µ + j + µ j )ku j k − 2 + µ+ j µ j ku j k−1 .

(6.76)

− + + 4 2 Recalling (6.60), and noting that µ − j µ j = j and µ j + µ j = 2α(1 + δ j ), we con− + clude from (6.76) that hhU j ,U j ii = 0, as claimed.

260

6

Examples

5.3. Having thus established the desired orthogonal decomposition, we proceed to show that A satisfies the spectral gap inequality (5.83), with respect to the equivalent norm ||| · ||| in X0 defined by the scalar product (6.75). For this, we first need to estimate the Lipschitz constant `F of F; recall that F(U) := (0, g(u)), and that we are assuming that g is globally Lipschitz continuous. Let P1 : X → X1 and P2 : X → X2 be the orthogonal projections corresponding to the decomposition X0 = X1 ⊕ X2 . P1 and P2 induce corresponding projections p1 and p2 in H1 and H−1 in a natural way. Recalling (6.71), (6.73), it follows that, for U = (u, v) ∈ X0 , 2

|||U||| = Φ(P1U, P1U) +Ψ (P2U, P2U) ≥ 2αkp1 uk2 + (αδ − 1)kp2 uk2 ≥ (αδ − 1) kpN uk2 + kp2 uk2 = (αδ − 1)kuk2 . (6.77) Given then U = (u, u) ˜ and V = (v, v) ˜ ∈ X , we compute |||F(U) − F(V )||| = kg(u) − g(v)k−1 ≤ ` k∇u − ∇vk ≤ √

` |||U −V ||| , αδ − 1

the last step following form (6.77) (recall that we are assuming αδ > 1). Thus, `F ≤ √

` . αδ − 1

(6.78)

5.4. We are now ready to conclude. By (6.78), the spectral gap inequality is satisfied if − µN+1 − µN− > √

` . αδ − 1

(6.79)

Recalling (6.63), we compute that − µN+1 − µN− = αδ (2N + 1) +

p p R(N) − R(N + 1) .

We shall prove below that p p p R(N) − R(N + 1) + (2N + 1) α 2 δ 2 − 1 = 0 ; lim N →+∞

(6.80)

(6.81)

assuming this for the moment, we can determine N1 > 0 such that for all N ≥ N1 , (6.62) holds. Then, from (6.80) we deduce that if N ≥ N1 , p − µN+1 (6.82) − µN− ≥ (2N + 1)(αδ − α 2 δ 2 − 1) − 1 . This means that if (6.81) holds and N ≥ N1 satisfies (6.61), the spectral gap inequality (6.79) follows from (6.80) and (6.82). To prove (6.81), setting R1 (N) := 1 +

2α 2 δ α2 + , (α 2 δ 2 − 1)N 2 (α 2 δ 2 − 1)N 4

6.2

261

Beam and von Kármán Equation

we compute that p p p R(N) − R(N + 1) + (2N + 1) α 2 δ 2 − 1 p p p = α 2 δ 2 − 1 (N + 1)2 1 − R1 (N + 1) − N 2 1 − R1 (N) . (6.83) We easily see that p lim N 2 1 − R1 (N) = −

N →+∞

α 2δ α 2δ 2 − 1

;

consequently, (6.83) yields (6.81). This concludes the proof of theorem 6.9 if αδ > 1.

6.2 Beam and von Kármán Equation In this section we consider the generalized BEAM EQUATION Z 2 2 |∇u| dx − β ∆u + f , εutt + ut + ∆ u =

(6.84)

Ω

with ε > 0 and β ∈ R. Equation (6.84) describes the displacement of a solid beam, filling a bounded domain Ω ⊆ RN , subject to an external load force f . As in the Cahn-Hilliard equations (6.1), the principal part of equation (6.84) is of the fourth order in the space variables; thus, the equation can be regarded as semilinear. Note, however, that the nonlinearity is nonlocal in character, due to the coefficient k∇uk2 at the right side of (6.84). In the sequel, we assume for simplicity that 0 < ε ≤ 1. Most of the material we present is taken from Eden-Milani, [EM93]; for more information on general types of beam equations, we refer to the references therein, and, in particular, to Ball, [Bal73].

6.2.1 Functional Framework and Notations We assume that Ω has a Lipschitz continuous boundary ∂ Ω , on which we impose the so-called “hinged” boundary conditions u|∂ Ω = 0 , ∆u|∂ Ω = 0 .

(6.85)

Finally, we supplement (6.84) with the initial conditions u(0, ·) = u0 , ut (0, ·) = u1 , and refer to the initial-boundary value problem (6.84)+(6.85)+(6.86) as

(6.86)

262

6

Examples

problem

(BE) .

We set H := L2 (Ω ), with the usual norm k · k and scalar product h · , · i, and define V := H2 (Ω ) ∩ H10 (Ω ) , D := {u ∈ H4 (Ω ) : u, ∆u ∈ H10 (Ω )} . Then, V ,→ H ,→ V 0 is a Gelfand triple, with compact injection V ,→ H. We also consider the product spaces X0 := V × H , X1 := D × V , on which we define the functions 1 kuk2 + k∆uk2 , E0 (u, v) := εkvk2 + hu, vi + 2ε

Φ0 (u, v) := E0 (u, v) + k∇uk4 − β k∇uk2 . By Schwarz’ inequality, we immediately see that E0 is the square of an equivalent norm in X0 . Likewise, we define the following functions on the space X1 : E1 (u, v) := εk∆vk2 + εh∆u, ∆vi + 12 k∆uk2 + k∆2 uk2 , Φ1 (u, v) := E1 (u, v) + k∇uk2 k∇∆uk2 − β k∇∆uk2 − 2h f (t, ·), ∆2 ui . Again, we easily check that, since we are assuming ε ∈ ]0, 1], E1 is the square of an equivalent norm in X1 .

6.2.2 The Beam Equation Semiflow As for equation (6.9) of section 3.1, we will show that problem (BE) generates a semiflow S, both in X0 and in X1 . At first, we recall the following global existence, uniqueness and regularity result. THEOREM 6.11 For all β ∈ R, (u0 , u1 ) ∈ X0 and f ∈ C([0, +∞[; H), there exists a unique u ∈ C([0, +∞[; V) ∩ C1 ([0, +∞[; H) , which is a weak solution of problem (BE) (i.e. with (6.84) satisfied in V 0 , almost everywhere in t). If in addition (u0 , u1 ) ∈ X1 and f ∈ C1 ([0, +∞[; H), then u ∈ C([0, +∞[; D) ∩ C1 ([0, +∞[; V) ∩ C2 ([0, +∞[; H) . PROOF See e.g. Ball, [Bal73]. Theorem 6.11 allows us to define the solution operator S = (S(t))t ≥0 in X0 , associated to problem (BE). When f is independent of t, S is a semigroup; to show that S is also a semiflow, it is sufficient to prove the following

6.2

Beam and von Kármán Equation

263

PROPOSITION 6.12 For each t ≥ 0, S(t) is locally Lipschitz continuous in X0 . PROOF Assume that u and u¯ are two solutions of (6.84), and let z := u − u. ¯ Then, z satisfies the equation εztt + zt + ∆2 z − k∇uk2 − β ∆z = k∇uk2 − k∇uk ¯ 2 ∆u¯ =: g . Multiplying this by 2zt and

1 ε

(6.87)

z in H, and adding the resulting identities, we obtain

d E0 (z, zt ) + k∇uk2 k∇zk2 + kzt k2 + ε1 k∆zk2 + ε1 k∇uk2 k∇zk2 dt = 2h∇u, ∇ut i k∇zk2 + 2β h∇z, ∇zt i + β ε1 k∇zk2 + ε1 hg, 2εzt + zi .

(6.88)

We now use integration by parts, to write h∇u, ∇ut i = h−∆u, ut i , k∇zk2 = h−∆z, zi , and note that k∇uk2 − k∇uk ¯ 2 = −h∆u + ∆u, ¯ u − ui ¯ . Consequently, we obtain from (6.88) d E0 (z, zt ) + k∇uk2 k∇zk2 + kzt k2 + ε1 k∆zk2 + ε1 k∇uk2 k∇zk2 dt ≤ 2k∆uk kut k k∆zk kzk + |β | k∆zk kzt k + |β | ε1 k∆zk kzk + ε1 k∆u + ∆uk ¯ k∆uk ¯ kzk (εkzt k + kzk) .

(6.89)

√ Since the map [0, +∞[ 3 t 7→ (u(t, ·), εut (t, ·)) ∈ V is locally bounded, and analogously for u, ¯ recalling that we are assuming that ε ≤ 1 we deduce from (6.89) that d E0 (z, zt ) + k∇uk2 k∇zk2 + kzt k2 + ε1 k∆zk2 + ε1 k∇uk2 k∇zk2 dt 1 ≤ 2ε k∆zk2 + 41 +Cε kzt k2 +C ε1 kzk2 ,

(6.90)

where C depends on |β |, u and u. ¯ Integrating (6.90) and applying Gronwall’s inequality (2.62), we can then easily conclude the proof of proposition 6.12.

6.2.3 Absorbing Sets We now show the existence of absorbing sets for the semiflow S generated by problem (BE). 1.

We first show that S admits a bounded, positively invariant absorbing set in X0 .

264

6

Examples

PROPOSITION 6.13 Assume that f ∈ Cb ([0, +∞[; H). There exists R0 > 0, dependent on ε, such that the set B0 := {(u, v) ∈ X0 : Φ0 (u, v) ≤ R20 }

(6.91)

is bounded, positively invariant and absorbing for the solution operator S generated by problem (BE) in X0 (that is, B0 absorbs all bounded sets of X0 ). PROOF To show that B0 is bounded, we see that, if (u, v) ∈ B0 , E0 (u, v) = Φ0 (u, v) + β k∇uk2 − k∇uk4 ≤ Φ0 (u, v) + 41 β 2 ≤ R20 + 41 β 2 .

(6.92)

To show that B0 is positively invariant and absorbing, we establish an exponential inequality on Φ0 (u, ut ). Multiplying equation (6.84) in H by 2ut and ε1 u, and adding the resulting identities, we obtain d Φ0 (u, ut ) + kut k2 + ε1 k∆uk2 + ε1 k∇uk2 − β k∇uk2 = ε1 h f , 2εut + ui . (6.93) dt Let λ1 denote the first eigenvalue of the operator (−∆)2 , relative to the boundary conditions (6.85), so that λ1 kuk2 ≤ k∆uk2 . Then, we can estimate the right side of (6.93) by 1 1 2k f k2 + 12 kut k2 + 2ελ k∆uk2 , k f k2 + 2ε 1

and, therefore, obtain from (6.93) d 1 Φ0 (u, ut ) + 12 kut k2 + 2ε k∆uk2 + ε1 k∇uk2 − β k∇uk2 ≤ ε1 C f , dt where the constant C f depends on supt ≥0 k f (t)k. We now set α := max{3, 2(λ1−1 + 1)} ,

(6.94)

(6.95)

and note that, since ε ≤ 1, (6.96) Φ0 (u, ut ) ≤ 23 εkut k2 + ( λ1ε + 1)k∆uk2 + k∇uk4 − β k∇uk2 1 1 1 ≤ α 21 kut k2 + 2ε k∆uk2 + 2ε k∇uk4 − ε1 β k∇uk2 + β ( αε − 1)k∇uk2 . Consequently, we deduce from (6.94) that d 1 k∇uk4 ≤ ε1 C f + ε1 |β (1 − αε)| · k∇uk2 Φ0 (u, ut ) + α1 Φ0 (u, ut ) + 2ε dt 1 1 ≤ ε1 C f + 2ε |β (1 − αε)|2 + 2ε k∇uk4 . Thus, Φ0 (u, ut ) satisfies the exponential inequality d 1 Φ0 (u, ut ) + α1 Φ0 (u, ut ) ≤ M1ε := ε1 C f + 2ε |β (1 − αε)|2 , dt so that we can conclude, as usual, the existence of a bounded, positively invariant absorbing set for S in X0 . Note that M1ε is unbounded as ε → 0.

6.2

265

Beam and von Kármán Equation

2. We now proceed to show that S also admits a bounded, positively invariant absorbing set in X1 . PROPOSITION 6.14 Assume that f ∈ C1b ([0, +∞[; H), and let B0 be the set defined in (6.91). There exists R1 > 0, dependent on ε, such that the set B1 := {(u, v) ∈ X1 : Φ1 (u, v) ≤ R21 } ∩ B0

(6.97)

is bounded, positively invariant and absorbing for the solution operator S generated by problem (BE) in X1 (that is, B0 absorbs all bounded sets of X1 ). PROOF To show that B1 is bounded, let (u, v) ∈ B1 , and set F0 := supt ≥0 k f (t)k. If k∇uk2 ≥ β , we have that E1 (u, v) ≤ Φ1 (u, v) + 2h f , ∆2 ui ≤ R21 + 2k f k2 + 21 k∆2 uk2 . Thus, since clearly k∆2 uk2 ≤ E1 (u, v) for all (u, v) ∈ X1 , 1 2 E1 (u, v) ≤

R21 + 2F02 .

(6.98)

If instead k∇uk2 ≤ β , then β > 0, and E1 (u, v) ≤ Φ1 (u, v) + β k∇∆uk2 + 2h f , ∆2 i .

(6.99)

By the Gagliardo-Nirenberg inequalities (see theorem A.70), and elliptic estimates similar to those of theorem A.77, we can estimate k∇∆uk ≤ Ck∆2 uk2/3 k∇uk1/3 ;

(6.100)

thus, we obtain from (6.99) that E1 (u, v) ≤ R21 +Cβ 5/3 k∆2 uk4/3 + 2k f k k∆2 uk ≤ R21 +C3 β 5 + 4k f k2 + 21 k∆2 uk . Consequently, 1 2 E1 (u, v) ≤

R21 +C3 β 5 + 4F02 .

(6.101)

Together with (6.98), (6.101) shows that B1 is bounded in X1 . To show that B1 is positively invariant and absorbing, we establish an exponential inequality on Φ1 (u, ut ). Multiplying equation (6.84) in H by 2∆2 ut and ∆2 u, and adding the resulting identities, we obtain d Φ1 (u, ut ) + (2 − ε)k∆ut k2 + k∆2 uk2 + k∇uk2 − β k∇∆uk2 − h f , ∆2 ui dt = 2h∇u, ∇ut ik∇∆uk − 2h ft , ∆2 ui

266

6

Examples

= 2h∆u, ut ih∆2 u, ∆ui − 2h ft , ∆2 ui ≤ 2k∆uk2 kut k k∆2 uk + 2k ft k k∆2 uk ≤ 8k∆uk4 kut k2 + 8k ft k2 + 14 k∆2 uk2 . Therefore, setting F1 := sup k ft (t)k and, for (u, v) ∈ X1 , t ≥0

Ψ1 (u, v) := k∆vk2 + 21 k∆2 uk2 + k∇uk2 − β k∇∆uk2 − h f , ∆2 ui , and recalling that ε ≤ 1, we deduce that d Φ1 (u, ut ) +Ψ1 (u, ut ) + 41 k∆2 uk2 ≤ 8F12 + 8k∆uk4 kut k2 . dt

(6.102)

We now easily verify that, with α defined in (6.95), Φ1 (u, v) ≤ αΨ1 (u, v) + (1 − α) k∇uk2 − β k∇∆uk2 + (α − 2)h f , ∆2 ui . Consequently, we obtain from (6.102) that d Φ1 (u, ut ) + α1 Φ1 (u, ut ) + 41 k∆2 uk2 ≤ 8F12 + 8k∆uk4 kut k2 dt + 1−α α k∇uk2 − β k∇∆uk2 + α α−2 k f k k∆2 uk .

(6.103)

Recalling that α > 1, we have then that, if k∇uk2 ≥ β , d Φ1 (u, ut ) + α1 Φ1 (u, ut ) ≤ 8F12 + 8k∆uk4 kut k2 + F02 . dt

(6.104)

If instead k∇uk2 ≤ β , resorting again to estimate (6.100) we proceed from (6.103) with d Φ1 (u, ut ) + α1 Φ1 (u, ut ) + 41 k∆2 uk2 dt ≤ 8F12 + 8k∆uk4 kut k2 +Cβ 5/3 k∆2 uk4/3 + k f k k∆2 uk ≤ 8F12 + 8k∆uk4 kut k2 +C1 β 5 + 2k f k2 + 14 k∆2 uk2 . Together with (6.104), this shows that, in either case, there exists a constant K > 0, depending on F0 , F1 and β , such that d Φ1 (u, ut ) + α1 Φ1 (u, ut ) ≤ K + 8k∆uk4 kut k2 . dt

(6.105)

In the sequel, we denote by Mr,ε , r ≥ 1, various positive constants, independent of u (but unbounded as ε → 0). Assume now that (u0 , u1 ) is in a bounded set of X1 . Then, (u0 , u1 ) is also in a bounded set of X0 , and since B0 is absorbing, there is T0 ≥ 0 such that Φ0 (u(t), ut (t)) ≤ R20 for all t ≥ T0 (with T0 = 0 if (u0 , u1 ) ∈ B1 , since B1 is positively invariant). We easily verify that, for each (u, v) ∈ X0 , k∆uk2 ≤ E0 (u, v) , kvk2 ≤ ε2 E0 (u, v) ;

(6.106)

6.2

Beam and von Kármán Equation

267

hence, recalling (6.92), we deduce that 8k∆uk4 kut k2 ≤

16 ε

R20 + 14 β

3

=: M2ε .

Inserting this into (6.105), we obtain the exponential inequality d Φ1 (u, ut ) + α1 Φ1 (u, ut ) ≤ K + M2ε =: M3ε , dt

(6.107)

from which we can conclude that if R21 > αM3ε , the set B1 defined in (6.97) is positively invariant and absorbing for S in X1 . Indeed, from (6.107) we obtain that, for t ≥ T0 , Φ1 (u(t), ut (t)) ≤ (Φ1 (u(T0 ), ut (T0 )) − αM3ε ) e−(t −T0 )/α + αM3ε . Consequently, we deduce that (u(t), ut (t)) ∈ B1 for all t ≥ T1 , with T1 > T0 defined by the identity (Φ1 (u(T0 ), ut (T0 )) − αM3ε ) e−(T1 −T0 )/α + αM3ε = R21 , if Φ1 (u(T0 ), ut (T0 )) > R21 , or T1 = T0 if instead Φ1 (u(T0 ), ut (T0 )) ≤ R21 . This concludes the proof of proposition 6.14; note that R1 depends on R0 , via M2ε .

6.2.4 The Global Attractor In this section we resort to theorem 2.56 of chapter 2 to show that the semiflow S generated by problem (BE) admits a global attractor A in X0 . More precisely, we resort to the α-contraction method described in section 3.4.5, to show that A = ω(B0 ), where B0 is the absorbing set determined in proposition 6.13. Note that ω(B0 ) is not empty, since it contains the stationary solutions of problem (BE). Recalling proposition 2.59, to apply theorem 2.56 it is sufficient to find an appropriate pseudometric δ on X0 , and a number t∗ > 0, such that condition (2.47) of chapter 3 holds, with T = S(t∗ ). For fixed τ > 0 we define in X0 × X0 the function Z τ 1/2 2 δτ ((u, v), (u, ¯ v)) ¯ := kP1 (S(t)(u, v)) − P1 (S(t)(u, ¯ v))k ¯ dt , (6.108) 0

where P1 is the projection from X0 onto V. With exactly the same proof of proposition 3.25, we have that, for each τ > 0, δτ is a pseudometric on X , precompact on B0 with respect to the norm of X0 defined by E0 . We proceed then to establish an estimate of the difference of two solutions of problem (BE), that allows us to apply proposition 2.59. PROPOSITION 6.15 There are positive constants γi , i = 1, 2, 3, depending on ε, β and R0 , but not on t, such that for all U0 := (u0 , u1 ), U 0 := (u¯0 , u¯1 ) ∈ B0 , and t ≥ 0, 2 (6.109) E0 (S(t)U0 − S(t)U 0 ) ≤ γ1 e−γ2 t/2 E0 (U0 −U 0 ) + γ3 δt (U0 ,U 0 ) ,

268

6

Examples

with δt defined in (6.108). PROOF We start from (6.89), which we rewrite as d E0 (z, zt ) + k∇uk2 − β k∇zk2 + kzt k2 + ε1 k∆zk2 + ε1 k∇uk2 − β k∇zk2 dt ≤ 2k∆uk kut k k∆zk kzk + ε1 k∆u + ∆uk ¯ k∆uk ¯ kzk (εkzt k + kzk) =: ρ1 . (6.110) Recalling (6.106), and that U0 , U 0 ∈ B0 , we have that k∆uk2 ≤ R20 + 14 β , εkut k2 ≤ 2 R20 + 14 β . Thus, the right side of (6.110) can be estimated by √ 1 ρ1 ≤ 2 2 √ R20 + 41 β k∆zk kzk + ε2 R20 + 14 β kzk (εkzt k + kzk) ε 1 k∆zk2 + 21 kzt k2 . ≤ C ε1 kzk2 + 4ε

Inserting this into (6.110), and setting Z := E0 (z, zt ) + k∇uk2 − β k∇zk2 , we obtain dZ + dt

2 2 1 1 1 2 kzt k + 2ε k∆zk + ε

1 k∆zk2 ≤ C ε1 kzk2 . (6.111) k∇uk2 − β k∇zk2 + 4ε

Acting as in (6.96), we easily see that Z≤α

2 2 1 1 1 2 kzt k + 2ε k∆zk + ε

k∇uk2 − β k∇zk2 + ε1 β (α − ε)k∇zk2 ;

consequently, we obtain from (6.111) that (for different C) dZ 1 1 1 + α Z + 4ε k∆zk2 ≤ C ε1 kzk2 +Ck∇zk2 ≤ C ε1 kzk2 + 4ε k∆zk2 . dt

(6.112)

From (6.112) we immediately obtain that, for all t ≥ 0, Z(t) ≤ Z(0)e−t/α +C ε1

Z t

kz(s)k2 ds ;

(6.113)

0

since we obviously have that, for suitable constant C, depending on R0 , and all t ≥ 0, E0 (z(t), zt (t)) ≤ Z(t) ≤ C E0 (z(t), zt (t)) , we obtain from (6.113) that, for all t ≥ 0, E0 (z(t), zt (t)) ≤ C E0 (z(0), zt (0))e−t/α +C ε1

Z t 0

kz(s)k2 ds ,

(6.114)

6.2

Beam and von Kármán Equation

269

from which (6.109) follows. The existence of a global attractor for S follows now theorem 2.56 and proposition 2.59. Indeed, choosing t∗ > 0 such that q := e−t∗ /2α < 1, from (6.109) we see that the operator T = S(t∗ ) and the pseudometric δt∗ satisfy condition (2.47) of proposition 2.59. Hence, T is an α-contraction, and the set A = ω(B0 ) is the desired attractor for S. REMARK 6.16 If ε is sufficiently small, it is possible to show, with techniques analogous to those of section 3.5, that the attractor A is bounded in X1 . For an alternative proof, see Eden-Milani, [EM93].

6.2.5 The Exponential Attractor In this section we show that the semiflow S also admits an exponential attractor E in X0 , which contains the global attractor A. THEOREM 6.17 In the same conditions of proposition 6.14, with ft ≡ 0, the semiflow S generated by problem (BE) admits an exponential attractor E in X0 . PROOF We apply theorem 4.5 of chapter 4. To this end, we consider the absorbing set B1 for S in X1 , determined in proposition 6.14; since the injection X1 ,→ X0 is compact (because so is the injection V ,→ H), B1 is compact in X0 . We propose then to show that S satisfies the discrete squeezing property (see definition 4.3), relative to B1 . We proceed almost exactly as in the proof of theorem 4.10 of section 4.4.2, of which we keep the same choices of N and XN , and denote by k · k0 the norm induced on X0 by E0 . Thus, we show that, given any t∗ > 0 and γ ∈ ]0, 12 [, there exists an integer N∗ , with the property that if u0 , u¯0 ∈ B1 are such that S(t∗ )u0 − S(t∗ )u¯0 ∈ / CN∗ (i.e. if (6.42) holds for the operator S(t∗ )), then (6.43) must hold. To this end, we first note that, since kzk2 ≤ 4εE0 (z, zt ), using Gronwall’s inequality we obtain from (6.114) that E0 (z(t), zt (t)) ≤ CE0 (z(0), zt (0))e4Ct .

(6.115)

Next, we apply the projection qN to the equation (6.87) satisfied by z, to obtain that the function q := qN z satisfies the equation εqtt + qt + ∆2 q − k∇uk2 − β ∆q = −h∆u + ∆u, ¯ zi∆qN u¯ . For (q, p) ∈ QN (X0 ), we define 1 k∇uk2 − β k∇qk2 . M(q, p) := εkpk2 + hp, qi + k∆qk2 + 2ε

(6.116)

270

6

Examples

Multiplying (6.116) in H by 2qt and ε1 q, and adding the resulting identities, we obtain d 1 1 M(q, qt ) + kqt k2 + 2ε hq, qt i + ε1 k∆qk2 + 2ε k∇uk2 − β k∇qk2 dt ¯ zi h∆qN u, ¯ 2εqt + qi = 2h−∆u, ut ik∇qk2 + ε1 h∆u + ∆u, 2 2 1 1 − 2ε hq, qt i − 2ε k∇uk − β k∇qk =: ρ2 . (6.117) We now proceed as in the proof of proposition 6.15: Recalling that U0 and U 0 ∈ B0 , that ε ≤ 1, and that, in analogy to (3.28) of proposition 3.6, kqk ≤ p

1

k∆qk ,

(6.118)

λN+1

we can estimate the right side of (6.117) by 2C0 2C0 ρ2 ≤ p k∆qk2 + p k∆uk ¯ kzk kqt k ελN+1 λN+1 C0 1 |β | + p kzk k∆qk + p k∆qk kqt k + p k∆qk2 ε λN+1 2ε λN+1 2ε λN+1 C0 C0 1 k∆qk2 + kzk2 + 12 kqt k2 + 4ε ≤ 2 k∆qk2 , (6.119) ε λN+1 ελN+1 where the constant C0 depends on R0 and β . Since λN → +∞ as N → +∞, we can choose N1 ∈ N so large that if N ≥ N1 , 4C0 ≤ ελN+1 . For such N, we obtain from (6.117) and (6.119), together with (6.115), that 4C0 d M(q, qt ) + 21 M(q, qt ) ≤ E0 (z(0), zt (0)) e4Ct . dt λN+1 Integrating this inequality, we obtain that, for all t ≥ 0, M(q(t), qt (t)) ≤ M(q(0), qt (0))e−t/2ε +

8εC0 e4Ct . (1 + 8Cε)λN+1

(6.120)

This estimate is the analogous of estimate (4.48) of section 4.4.2; it follows that, to conclude the proof of (6.43), it is sufficient to choose t∗ and N∗ so that the right side of (6.120) is (arbitrarily) small. Thus, given any η > 0, we first choose t∗ so large that M(q(0), qt (0))e−t∗ /2ε ≤ η, and then N∗ ≥ N1 so large that 8εC0 e4Ct∗ ≤ η . (1 + 8Cε)λN∗ +1 With these choices, we obtain from (6.120) that M(q(t∗ ), qt (t∗ )) ≤ 2η . Proceeding then as in the proof of theorem 4.10, we can then deduce that the discrete squeezing property holds, relative to the set B1 . By theorem 4.5, this is sufficient to conclude the proof of theorem 6.17.

6.2

271

Beam and von Kármán Equation

6.2.6 Inertial Manifold In this last section, we show that if the source term f in (6.84) has the special form f = pN f , for some N ∈ N, where pN = IH − qN is the projection on H considered in the last section, then the semiflow S admits a (trivial) inertial manifold. More precisely, we have THEOREM 6.18 Let N ∈ N and f ∈ H be such that f ∈ pN H. There exists M ≥ N such that the flat manifold M := PM X0 is an inertial manifold for S. PROOF Let u be a solution of problem (BE), corresponding to initial values U0 = (u0 , u1 ) in a bounded set G of X0 , and set U(t) := (u(t), ut (t)) = S(t)U0 . Fix M ≥ N, to be determined. Since ∂ (S(t)U0 , M) ≤ E0 (U(t) − PMU(t)) , setting q := qM u it is sufficient to show that E0 (U(t) − PMU(t)) = E0 (q(t), qt (t)) ≤ k1 e−k2 t ,

(6.121)

for suitable positive constants k1 , k2 , depending only on G. To this end, we see that, since qM f = 0 because M ≥ N, q satisfies the equation εqtt + qt + ∆2 q − k∇uk2 − β ∆q = 0 . This equation is similar to (6.116); acting as in (6.117), we arrive at the estimate d 1 1 M(q, qt ) + 21 kqt k2 + 2ε k∆qk2 M(q, qt ) + 2ε dt 1 1 = 2h−∆u, ut ik∇qk2 − 2ε hq, qt i − 2ε k∇uk2 − β k∇qk2 ! 1 1 1 + + p k∆qk2 + 12 kqt k2 . (6.122) ≤ C0 p ελM+1 ε 2 λM+1 ε λM+1 Taking M ≥ N so large that C0

1 p

ελM+1

1 1 + 2 + p ε λM+1 ε λM+1

! ≤

1 , 2ε

we obtain from (6.122) that d 1 M(q, qt ) ≤ 0 . M(q, qt ) + 2ε dt This estimate yields the exponential decay of M(q, qt ); in turn, because of (6.118), this yields the exponential decay of E0 (q, qt ). Thus, we can deduce that (6.121) holds, and complete the proof of theorem 6.18.

272

6

Examples

6.2.7 von Kármán Equations We conclude this section with the remark that we can proceed with almost exactly the same techniques we have used for problem (BE), to study the long-time behavior of weak solutions to the equation p utt + δ ρut + ∂x4 u + 1 − κkux k2 − σ hux , uxt i uxx + α∂x4 ut + ρux = 0 , (6.123) in one-dimension of space (e.g., with Ω = ]0, 1[). This equation represents a onedimensional version of the so-called VON K ÁRMÁN equations for a thin plate; more precisely, it describes the displacement of a thin elastic plate subject to an axial force load and to the flow of a fluid along its surface. In (6.123), the positive parameters α, σ , δ represent various damping parameters associated to the plate and the fluid; κ is a measure of the elastic properties of the plate, and ρ is the flow rate of the fluid flowing along its surface. We refer to Guckenheimer-Holmes, [GH83, sct. VII.7.6], for a detailed study of various IBV problems associated to (6.123), and for the consequent definition of the associated semiflow. In Eden-Milani, [EM93], we briefly outline the argument leading to the existence of a global and an exponential attractor for this semiflow, in a phase space analogous to the space X0 considered for problem (BE).

6.3 Navier-Stokes Equations In this section we consider the so-called NAVIER -S TOKES equations in two dimensions of space. These equations describe the motion of a two-dimensional viscous, incompressible fluid in a bounded set Ω ⊂ R2 . For more information on the NavierStokes equations, also in three space dimensions, we refer e.g. to Lions, [Lio69, sct. 1.6], Temam, [Tem83], Constantin-Foias, [CF88], and Sell-You, [SY02, ch. 6]. For j = 1, 2, we set ∂ j := ∂ /∂ x j . Given two smooth vector fields ~a = (a1 , a2 ) and ~b = (b1 , b2 ), we define a third field (~a · ∇)~b by (~a · ∇)~b := (a1 ∂1 b1 + a2 ∂2 b1 , a1 ∂1 b2 + a2 ∂2 b2 ) .

(6.124)

6.3.1 The Equations and their Functional Framework Denoting by ~u the velocity of the fluid, and by p its pressure, the Navier-Stokes equations we consider have the form ρ0 (~ut + (~u · ∇)~u) − ν∆~u = ~f − ∇p , div~u = 0 ,

(6.125) (6.126)

where ρ0 > 0 is the density of the fluid (which we assume to be constant; in the sequel, we take ρ0 = 1), ν > 0 is its kinematic viscosity, and ~f is a measure of the external forces applied to the fluid. ν is proportional to the reciprocal of the Reynolds

6.3

Navier-Stokes Equations

273

number; in many fluids, ν is small. Equation (6.126) translates the requirement that the fluid be incompressible. We supplement (6.125) and (6.126) with the initial and boundary conditions ~u(0, x) = ~u0 (x) , x∈Ω, ~u(t, x) = 0 , (t, x) ∈ [0, +∞[ ×∂ Ω .

(6.127)

The homogeneous boundary condition in (6.127) means that we assume the fluid to be at rest at the boundary of Ω ; other type of boundary conditions can be considered, such as space-periodic ones, or so-called “nonpenetrating” boundary conditions, of the form ~n ·~u = 0 , ~n × curl~u = 0 , where ~n is the outward unit normal to the boundary ∂ Ω . We refer to the evolution problem (6.125) + (6.126) + (6.127) as problem

(NS) .

The function spaces in which we consider problem (NS) are suitable subspaces of the spaces H(div, Ω ) and H(curl, Ω ) introduced in section A.7, to which we refer. In particular, we adopt its convention A.83, whereby if X is a space of scalar functions on Ω , such as L2 (Ω ), then, with abuse of notation, we write ~u ∈ X to mean that the components of ~u are in X . Thus, we rely on the context to know when the notation X denotes a space of scalar valued or of vector valued functions (i.e. when X is actually an abbreviation for the space X × X ). Also, when there is no danger of confusion, we denote vectors simply by u, instead of ~u. Finally, we denote as usual the norm and scalar product in L2 (Ω ) by k · k and h · , · i, and by | · | p the norm in L p (Ω ), 1 ≤ p ≤ +∞. Setting then V := {~u ∈ H10 (Ω ) : div~u = 0} , H := H0 (div, Ω ) = {~u ∈ L2 (Ω ) : div~u = 0} , theorem A.85 allows us to consider V ,→ H ,→ V 0 as a Gelfand triple. Next, recalling (6.124), we define a trilinear form b on V, by 2

b(u, v, w) := h(u · ∇)v, wi =

∑

Z

ui (∂i v j )w j dx .

(6.128)

i, j=1 Ω

The following proposition describes the main properties of b that we need in the sequel. PROPOSITION 6.19 The trilinear form b defined in (6.128) is a continuous map from V × V × V to R, which satisfies the estimate |b(u, v, w)| ≤ |u|4 |∇v|2 |w|4 ,

(6.129)

274

6

Examples

for all u, v and w ∈ V. Moreover, for all u, v ∈ V, b(u, u, v) = −b(u, v, u) .

(6.130)

b(u, v, v) = 0 , b(u, u, u) = 0 .

(6.131)

In particular,

PROOF Estimate (6.129) follows from theorem A.58 (recall that, since N = 2, V ,→ H10 (Ω ) ,→ L p (Ω ) for all p ∈ [2, +∞[). To prove identity (6.130), we use the integration by parts formula Z

Z

(~n · ~f )g ds =

(div ~f )g ds +

Z

~f · ∇g dx ,

(6.132)

Ω

Ω

∂Ω

where ~n is the outward unit normal to ∂ Ω ; the validity of this formula for functions in V is justified in section A.7. Indeed, we easily compute that Z

b(u, u, v) + b(u, v, u) =

~u · ∇(~u ·~v) dx ;

Ω

hence, by (6.132) we have that b(u, u, v) + b(u, v, u) = −

Z

(div~u)(~u ·~v) dx +

Z

(~n ·~u)(~u ·~v) ds = 0 ,

∂Ω

Ω

keeping in mind that u ∈ V. PROPOSITION 6.20 Let T > 0, and u ∈ L2 (0, T ; V) ∩ L∞ (0, T ; H). Then, (u · ∇)u ∈ L2 (0, T ; V 0 ). PROOF Let v ∈ L2 (0, T ; V). By the Gagliardo-Nirenberg inequality |u|4 ≤ Ck∇uk1/2 kuk1/2 ,

(6.133)

which holds because u vanishes at ∂ Ω (see theorem A.70), recalling (6.130) and (6.129), we can estimate Z TZ 0

|(u · ∇)u · v| dx dt =

Z T

|b(u, u, v)| dt =

Z T

≤

Z T 0

|b(u, v, u)| dt

0

0

Ω

|u|24 k∇vk dt ≤ C

Z T

k∇uk kuk k∇vk dt

0

≤ CkukL∞ (0,T ;H) kukL2 (0,T ;V ) kvkL2 (0,T ;V ) , from which the conclusion follows.

6.3

Navier-Stokes Equations

275

6.3.2 The 2-Dimensional Navier-Stokes Semiflow Global existence of weak solutions to problem (NS) is provided by the following result, a proof of which can be found e.g. in Lions, [Lio69, sct. 1.6] (see also Temam, [Tem83]). THEOREM 6.21 Let T > 0. For all f ∈ L2 (0, T ; V 0 ) and u0 ∈ H, there exists a unique u ∈ L2 (0, T ; V), with ut ∈ L2 (0, T ; V 0 ), which is a weak solution of problem (NS), in the sense that u(0, ·) = u0 (this makes sense because, by theorem A.80, u ∈ C([0, T ]; H)), and for all v ∈ V, hut + (u · ∇)u − ν∆u − f , viV 0 ×V = 0 ,

(6.134)

almost everywhere in t ∈ [0, T ]. If in addition f ∈ L2 (0, T ; H), ft ∈ L2 (0, T ; V 0 ), then u ∈ C([τ, T ]; V) ∩ L2 (τ, T ; H2 (Ω ) ∩ V)

(6.135)

for all τ ∈ ]0, T [. If u0 ∈ V, we can take τ = 0 in (6.135). REMARK 6.22 In the weak formulation (6.134), the unknown p is not present. This is because, by (6.132), h∇p, viV 0 ×V = 0 for all v ∈ V. On the other hand, once a weak solution u of problem (NS) has been found, we can formally determine p in the following way (this procedure can be justified rigorously by the results of section A.7). By proposition 6.20, we know that w := ut + (u · ∇)u − ν∆u − f ∈ V 0 for almost all t ∈ [0, T ]. Let now ψ ∈ C∞ 0 (Ω ). Then, curl ψ ∈ V, so that (6.134) implies that hcurl w, ψiV 0 ×V = hw, curl ψiV 0 ×V = 0 . The arbitrariness of ψ implies then that curl w = 0; thus, there is a scalar function p such that w = −∇p, as desired. Theorem 6.21 allows us to define the solution operator S = (S(t))t ≥0 in H, associated to problem (NS). When f is independent of t, S is a semigroup; to show that S is also a semiflow, it is sufficient to prove the following PROPOSITION 6.23 For each t ≥ 0, S(t) is locally Lipschitz continuous in H. PROOF Assume that u and u¯ are two solutions of (6.134), and let z := u − u. ¯ Then, z satisfies the equations hzt − ν∆z, viV 0 ×V + b(u, z, v) + b(z, u, ¯ v) = 0 ,

(6.136)

276

6

Examples

for all v ∈ V. Take now v = z in (6.136) (which is legitimate), and integrate in [0,t]: recalling (6.131) and (6.133), we obtain 2 1 2 kz(t)k + ν

Z t 0

k∇zk2 dt = 12 kz(0)k2 −

Z t

b(z, u, ¯ z) dt 0

Z t

≤ 21 kz(0)k2 +C ≤ 21 kz(0)k2 +C ≤ 21 kz(0)k2 + ν

0

Z t 0 Z t

¯ dt |z|24 k∇uk kzk k∇zk k∇uk ¯ dt k∇zk2 dt +C

Z t

kzk2 k∇uk ¯ 2 dt .

0

0

Consequently, we obtain that kz(t)k2 ≤ kz(0)k2 +C

Z t

kzk2 k∇uk ¯ 2 dt ;

0

recalling that u¯ ∈ L2 (0, T ; V), the conclusion follows by Gronwall’s inequality (2.62).

We remark that the validity of theorem 6.21 and proposition 6.23 is strictly limited to the two-dimensional case (N = 2). For example, as the proof of proposition 6.23 shows, if N = 3 the question of uniqueness of weak solutions is open.

6.3.3 Absorbing Sets and Attractor We now show the existence of a bounded, positively invariant absorbing set for S, first in H and then in V. As a consequence of the asymptotic smoothness of the semiflow, we deduce that S also admits a global attractor in H. PROPOSITION 6.24 Assume that f ∈ L∞ (0, +∞; V 0 ). There exists a positively invariant, absorbing ball B ⊆ H for the solution operator S generated by problem (NS) (that is, B absorbs all bounded sets of H). If in addition f ∈ L∞ (0, +∞; H), S also admits a positively invariant absorbing ball B1 in V (that is, B1 absorbs all bounded sets of V). PROOF In the sequel, we set F0 := sup k f (t)k2V 0 , F1 := sup k f (t)k2H . t ≥0

t ≥0

1. Multiplying equation (6.125) in H by 2u, and recalling the second of (6.131), we obtain that d kuk2 + 2νk∇uk2 = 2h f , uiV 0 ×V ≤ Ck f k2V 0 + νk∇uk2 . dt

(6.137)

6.3

Navier-Stokes Equations

277

By Poincaré’s inequality (3.16), we deduce from (6.137) the exponential inequality d kuk2 + νλ1 kuk2 ≤ CF0 . dt The existence of a positively invariant absorbing ball for S in H follows then by proposition 2.64. 2. To prove the existence of an absorbing ball in V, we formally multiply equation (6.125) in H by −2∆u. Since we do not know that −∆u ∈ H, this procedure is formal; to fully justify it, we should consider, as usual, the Galerkin approximations of u, constructed on subspaces of the eigenvalues of −∆ (with respect to the homogeneous Dirichlet boundary conditions). We obtain that d k∇uk2 + 2νk∆uk2 + 2b(u, u, −∆u) = 2h f , −∆ui dt ≤ Ck f k2 + 21 νk∆uk2 .

(6.138)

Recalling (6.130), and resorting to the Gagliardo-Nirenberg and elliptic estimates |u|4 ≤ Ck∆2 uk1/4 kuk3/4 ≤ Ck∆uk1/4 kuk3/4 (see theorems A.70 and A.77), we estimate b(u, u, −∆u) = b(u, ∆u, u) ≤ C|u|24 k∆uk ≤ Ckuk3/2 k∆uk3/2 ≤ Ckuk6 + 12 νk∆uk2 . Inserting this into (6.138), we obtain that d k∇uk2 + νk∆uk2 ≤ CF1 +Ckuk6 . dt

(6.139)

Since the first part of this proof provides an estimate on kuk independent of t ≥ 0, recalling the second Poincaré inequality (3.17) we finally deduce from (6.139) the exponential inequality d k∇uk2 + νλ1 k∇uk2 ≤ CF1 +C1 . dt

(6.140)

The existence of a positively invariant absorbing ball for S in V follows then again by proposition 2.64. This concludes the proof of proposition 6.24. For future reference, we remark that from (6.139) we also deduce that, for all t, s ∈ [0, +∞[, with s < t, Z t

ν

k∆uk2 dt ≤ k∇u(s)k2 +C(t − s)

s

and, since (6.140) implies that the function s 7→ k∇u(s)k is bounded, we conclude that there is C > 0, independent of s and t, such that Z t

ν s

k∆uk2 dt ≤ C(1 + t − s) .

(6.141)

278

6

Examples

We can now deduce the existence of a global attractor for the semiflow S in H. THEOREM 6.25 Let f (t, ·) ≡ f ∈ H. The semiflow S generated by problem (NS) admits a compact, global attractor A in H. PROOF It is sufficient to apply theorem 2.46 of chapter 2. Indeed, the last claim of theorem 6.21 implies that the semiflow S is uniformly compact for large t; hence, the desired attractor is the set A = ω(B), where B is the absorbing ball B constructed in proposition 6.24.

6.3.4 The Exponential Attractor In this section we show that the semiflow S also admits an exponential attractor E in H, which contains the global attractor A. THEOREM 6.26 In the same conditions of theorem 6.25, the semiflow S generated by problem (NS) admits an exponential attractor E in H. PROOF We apply theorem 4.5 of chapter 4. To this end, we consider the absorbing ball B1 for S in V, determined in proposition 6.24: since the injection V ,→ H is compact (because the map V ∈ u 7→ div u ∈ H is continuous), B1 is compact in H. We propose then to show that S satisfies the discrete squeezing property (see definition 4.3), relative to B1 . We proceed almost exactly as in the proof of theorem 4.7 of section 4.3.2, of which we keep the same choices of N and XN . In particular, we show that, given any t∗ > 0 and γ ∈ ]0, 21 [, there exists an integer N∗ , with the property that if u0 , v0 ∈ B1 are such that S(t∗ )u0 − S(t∗ )v0 ∈ / CN∗ , i.e. if kPN∗ (S(t∗ )u0 − S(t∗ )v0 )k < kQN∗ (S(t∗ )u0 − S(t∗ )v0 )k

(6.142)

(that is, if (4.13) holds for the operator S(t∗ )), then (4.11) must hold, i.e. kS(t∗ )u0 − S(t∗ )v0 k ≤ γku0 − v0 k .

(6.143)

Thus, we follow the evolution of the quotient norm Λ introduced in (3.58), i.e. Λ (t) :=

k∇z(t)k2 , kz(t)k2

where z(t) := S(t)u0 − S(t)v0 is the difference of the two solutions of problem (NS), with initial data u0 and v0 . Acting as in the proof of (3.61), we arrive at the identity 1 1 dΛ = h−∆z − Λ z, ν(∆z + Λ z) − (u · ∇)z − (z · ∇v)i . 2 dt kzk2

(6.144)

6.3 Setting w :=

z

k zk ,

Navier-Stokes Equations

279

we deduce from (6.144) that

dΛ = −2νk∆w + Λ wk2 + 2νh∆w + Λ w, (u · ∇)w + (w · ∇v)i dt ≤ ν k(u · ∇)wk2 + k(w · ∇)uk2 =: ν(R1 + R2 ) .

(6.145)

We first have that R1 ≤ C|u|2∞ k∇wk2 ;

(6.146)

thus, resorting to the Gagliardo-Nirenberg and elliptic inequalities |u|∞ ≤ Ck∆2 uk1/2 kuk1/2 ≤ Ck∆uk1/2 kuk1/2 (see theorems A.70 and A.77), and recalling that u ∈ B1 , so that its norm in V , and therefore in H, is uniformly bounded in t, we deduce from (6.146) that R1 ≤ C1 k∆ukΛ .

(6.147)

Similarly, since kwk = 1 and v ∈ B1 , recalling (6.133) we can estimate √ R2 ≤ C|w|24 |∇v|24 ≤ Ck∇wk kwk k∆vk k∇vk ≤ Ck∆vk Λ .

(6.148)

Inserting (6.147) and (6.148) into (6.145), we obtain that (for different constants C) dΛ ≤ Ck∆vk2 +C(1 + k∆uk)Λ . dt Integrating this inequality for 0 < s < t yields Λ (t) ≤ Λ (s) +C

Z t s

k∆vk2 dθ +C

Z t

(1 + k∆uk)Λ (s) dθ .

s

By Gronwall’s inequality, and recalling estimate (6.141), which also holds for v, we have then Zt Z t 2 Λ (t) ≤ Λ (s) +C k∆vk dθ exp C (1 + k∆uk) dθ s s √ √ ≤ (Λ (s) +C(1 + t − s)) exp C (t − s + t − s 1 + t − s) ≤ (Λ (s) +C(1 + t − s)) eC(1+t −s) , from which we obtain that Λ (s) ≥ e−C(1+t −s)Λ (t) −C(1 + t − s) . Integrating this inequality with respect to s in the interval [0,t], we finally deduce the estimate Z t 0

Λ (s) ds ≥ C1Λ (t)(1 − e−Ct ) −C 1 + t + 21 t 2 ,

(6.149)

280

6

Examples

with C1 := (CeC )−1 . Our next step is to recall (6.136) (with u¯ replaced by v): Since v ∈ B1 , which is bounded in V , we easily obtain that, as in (4.32), Z t kz(t)k2 ≤ kz(0)k2 exp C2t − ν Λ (s) ds , 0

where C2 depends on the bound on k∇vk provided by B1 . Consequently, setting ϕ(t) := C2t +C 1 + t + 21 t 2 , we deduce from (6.149) that kz(t)k2 ≤ kz(0)k2 exp ϕ(t) −C1Λ (t)(1 − e−Ct ) .

(6.150)

Let now t∗ > 0 be such that (6.142) holds: then, as in (4.29), Λ (t∗ ) ≥ 12 λN+1 . Consequently, (6.150) yields that kz(t∗ )k2 ≤ kz(0)k2 exp ϕ(t∗ ) −C1 λN+1 (1 − e−Ct∗ ) . Given then γ ∈ ]0, 21 [, we can make ϕ(t∗ ) −CλN+1 (1 − e−ct∗ ) ≤ 2 ln γ by choosing N so large that ϕ(t∗ ) − 2 ln γ ≤ CλN+1 (1 − e−ct∗ ) . Thus, (6.143) holds. The rest of the proof of theorem 6.26 proceeds then as that of theorem 4.7. We conclude this section with the remark that, in contrast to theorems 6.25 and 6.26, the existence of an inertial manifold for the semiflow generated by the NavierStokes equations is an open question, even in two dimensions of space, and for different types of boundary conditions, such as periodic ones.

6.4 Maxwell’s Equations In this section we consider a model for the quasi-stationary M AXWELL’ S equations, which describe the evolution of the electromagnetic fields and inductions in a ferromagnetic medium. This situation is characterized by a nonlinear dependence between the magnetic field and induction, and by the fact that the displacement currents are negligible with respect to the eddy ones. The former feature gives rise to a quasilinear system, while the latter allows us to consider a reduced problem, which is of parabolic type. For more information on Maxwell’s equations, we refer e.g. to Duvaut-Lions, [DL69, ch. 7].

6.4

Maxwell’s Equations

281

6.4.1 The Equations and their Functional Framework 1. The complete system of Maxwell’s equations is the first order linear system, essentially derived from the so-called A MPÈRE ’ S theorem and FARADAY ’ S LAW, Dt − curl H = G − J , Bt + curl E = 0 , div D = ρ , div B = 0 ,

(6.151) (6.152) (6.153) (6.154)

where E and H denote, respectively, the electric and the magnetic fields, and D, B, the corresponding inductions. The vector functions G and J in (6.151) represent, respectively, an external source, and the so-called eddy currents, while the scalar function ρ in (6.153) is a measure of the total electric charge. We refer to equations (6.151), ..., (6.154) collectively as system

(ME) .

This system is considered in a bounded domain Ω ⊂ R3 , with a Lipschitz boundary ∂ Ω . We supplement system (ME) with the initial conditions B(0, ·) = D0 , B(0, ·) = B0 ,

(6.155)

and impose the boundary conditions ν ×E = 0, ν ·B = 0,

(6.156)

where ν is the unit outward normal to ∂ Ω . The first of conditions (6.156) translates the assumption that the boundary of Ω be a so-called “perfect” conductor (i.e., an ideal metal). As we remark below, the second boundary condition in (6.156) is in general redundant, since it is a consequence of the first condition, and of equation (6.152). REMARK 6.27 The second Maxwell’s equation (6.152) implies that div Bt = 0 , ν · Bt = 0 (for the latter, see e.g. proposition A.89). Thus, if we assume, as it is customary, that the initial value B0 satisfies the conditions div B0 = 0 , ν · B0 = 0 ,

(6.157)

both equation (6.154) and the second boundary condition in (6.156) are a consequence of (6.152). System (ME) consists of eight conditions on twelve unknowns (the components of D, E, B and H). To make (ME) a determined system, we assume the constituent relations D = εE , H ∈ ζ (B) ,

282

6

Examples

where ε is, for simplicity, a positive constant (known as the dielectric constant, which measures the effects of the displacement currents), and ζ is a monotone map, in general multivalued because of the presence of hysteresis. Here, again for simplicity, we consider an idealized model, in which the hysteresis phenomena are neglected; thus, we assume that ζ is a monotone function. Finally, we assume that the eddy currents are everywhere present, and caused entirely by the conductivity of the medium; that is, that J = σE , where σ is, for simplicity, a positive constant measuring conductivity. With these assumptions, Maxwell’s equations (ME) can be written as εEt − curl ζ (B) = G − σ E , Bt + curl E = 0 , div E = ε1 ρ , div B = 0 ,

(6.158) (6.159) (6.160)

which we again refer to as system (ME) . Again, equation (6.160) is redundant if (6.157) holds. To make the system determined, we further assume that G and ρ satisfy the compatibility condition ρt + ε1 σ ρ = div G , which is derived by taking the divergence of (6.158). In ferromagnetic media, the effect of displacement currents is usually negligible in comparison to those of the eddy currents; that is, ε σ . It is then common, in applications, to neglect the term εEt in equation (6.158), and to consider instead the reduced equations σ E − curl ζ (B) = G , Bt + curl E = 0 . These equations are known as the refer to them as

QUASI - STATIONARY

(6.161) (6.162) Maxwell’s equations; we

system (QS) . We remark that, now, equation (6.159) loses sense, and the divergence of E is determined by equation (6.161), that is div E =

1 div G . σ

(6.163)

The first initial condition on D in (6.155) is also lost; indeed, system (QS) must be considered as a singular limit problem for the complete system (ME).

6.4

Maxwell’s Equations

283

2. We now transform the first order system (QS) into a formally parabolic evolution equation, by the introduction of suitable electromagnetic potentials. To describe this process, assume that system (QS) has a solution (B, E), with B(t, ·) ∈ H00 (div, Ω ) and E(t, ·) ∈ H(div, Ω ) for almost all t > 0 (see section A.7 for the definition of these and related spaces). By proposition A.91, we can then determine a vector function A(t, ·) ∈ H0 (curl, Ω ) ∩ H0 (div, Ω ), such that    curl A = B , (6.164) div A = 0 ,   ν ×A = 0. Then, equation (6.162) implies that At + E ∈ H0 (curl, Ω ) ∩ H(div, Ω ) for almost all t; hence, by theorem A.93, there is ϕ ∈ H10 (Ω ), such that At + E = −∇ϕ .

(6.165)

Recalling (6.163) and the second of the equations in (6.164), we see ϕ is determined as the solution of the boundary value problem   −∆ϕ = 1 div G , σ (6.166)  ϕ = 0 . ∂Ω

Replacing the expressions of B and E obtained in (6.164) and (6.165) into system (QS), we finally obtain the IBV problem    σ At + curl ζ (curl A) = −G − σ ∇ϕ =: F , (6.167) A(0, ·) = A0 ,   ν ×A = 0, where the initial value A0 is determined from B0 by means of the boundary value problem (6.164), at t = 0. REMARK 6.28 The equation in problem (6.167) is quasilinear, and has the same form as equation (3.11). As such, it does not immediately fit within the framework of the semilinear equations we have considered in chapter 3; however, if the function ζ is monotone, problem (6.167) is of parabolic type, and a suitable weak solution theory can be established, by means of classical results on evolution equations with monotone operators (see e.g. Brezis, [Bre73]). In contrast, for the corresponding IBV problem for the complete system of Maxwell’s equation (i.e. when ε > 0), which is  εAtt + σ At + curl ζ (curl A) = −G − ε∇ϕt − σ ∇ϕ =: F ,     A(0, ·) = A0 ,  At (0, ·) = A1 := −∇ϕ(0, ·) − E0 ,    ν ×A = 0, an analogous weak solution theory is not yet available. This is the principal reason why we consider only the reduced problem (QS).

284

6

Examples

3. In the sequel, we take for simplicity σ = 1. In accord with (6.157), we assume that B0 ∈ H00 (div, Ω ), and that the source term G satisfies G ∈ L2 (0, +∞; L2 (Ω )) ∩ L∞ (0, +∞; L2 (Ω )) , Gt ∈ L2 (0, +∞; L2 (Ω )) . (6.168) Then, A0 ∈ H0 (curl, Ω ) ∩ H0 (div, Ω ), and the function ϕ defined in (6.166) is such that ϕ and ϕt ∈ L2 (0, +∞; H10 (Ω )). Therefore, the source F in (6.167) has the same regularity (6.168) as G. We set then H := H0 (div Ω ) , V := H0 (curl, Ω ) ∩ H0 (div, Ω ) . By theorem A.85, we have that V ,→ H ,→ V 0 is a Gelfand triple. Moreover, by Friedrichs’ inequality (A.80), we can choose in V the norm kukV := k curl uk . In fact, since div u = 0 if u ∈ V, we have that kuk ≤ λF k curl uk ,

(6.169)

for all u ∈ V, with λF independent of u. Finally, we also set D := {u ∈ V : curl ζ (curl u) ∈ H0 (div Ω )} .

(6.170)

As we have stated above, we assume that ζ is a monotone function. More precisely, we assume that ζ : R3 → R3 is a globally Lipschitz continuous function, and the derivative of a convex function Z ∈ C1,1 (R3 , R). Without loss of generality, we can choose Z so that Z(0) = 0. We also assume that the derivative ζ 0 (p), which is defined for almost all p ∈ R3 , is a uniformly strictly positive matrix, that is ∃ γ > 0 ∀ q ∈ R3 :

hζ 0 (p)q, qiR3 ≥ γ|q|2 ,

(6.171)

with γ independent of p. We have then the following estimates: PROPOSITION 6.29 1. Let ζ satisfy the assumptions stated above, and L be the Lipschitz constant of ζ . Then, for all u and v ∈ H, γkuk2 + hζ (0), ui ≤ hζ (u), ui ≤ Lkuk2 + hζ (0), ui , 0

2

(6.172)

hζ (u)v, vi ≥ γkvk .

(6.173)

γkuk2V ≤ hζ (curl u), curl ui ≤ Lkuk2V .

(6.174)

In particular, for all u ∈ V,

6.4

285

Maxwell’s Equations

2. Given Z as above, define a function N : L2 (Ω ) → [0, +∞[ by Z

Z(u(x)) dx .

N(u) := Ω

Then, for all u ∈ L2 (Ω ), 2 1 2 γkuk + hζ (0), ui

≤ N(u) ≤ 12 Lkuk2 + hζ (0), ui .

(6.175)

In particular, for all u ∈ V, 2 1 2 γkukV

≤ N(curl u) ≤ 12 Lkuk2V .

(6.176)

PROOF 1. Since hζ (u), ui = hζ (u) − ζ (0), u − 0i + hζ (0), ui , (6.172) is an immediate consequence of the Lipschitz continuity of ζ and of (6.171), which also implies (6.173). Then, (6.174) follows from (6.172), since hζ (0), curl ui = 0, as follows by integration by parts. 2. Since ζ = Z 0 , and Z(0) = 0, recalling (6.172) we compute that, for u ∈ L2 (Ω ), Z

N(u) =

(Z(u(x)) − Z(0)) dx =

Ω

Z 1

= 0

Z Z 1

1 r hζ (ru), rui dr

≥ γkuk2

ζ (r u(x)) · u(x) dr dx

0

Ω

Z 1

Z 1

r dr + 0

hζ (0), ui .

0

This implies the first half of (6.175); the second half follows similarly. Finally, (6.176) follows from (6.175). REMARK 6.30 Under the stated assumptions on ζ , it can be shown, by means of elliptic regularity results similar to theorem A.77, that if D is the space defined in (6.170), then D = H2 (Ω ) ∩ V.

6.4.2 The Quasi-Stationary Maxwell Semiflow 1. The following theorem provides the global existence and regularity of weak solutions to problem (6.167). THEOREM 6.31 Let T > 0. For all F ∈ L2 (0, T ; V 0 ) and A0 ∈ H, there exists a unique A ∈ C([0, T ]; H) ∩ L2 (0, T ; V), with At ∈ L2 (0, T ; V 0 ), which is a weak solution of problem (6.167), in the sense that A(0, ·) = A0 (this makes sense because, by theorem A.80, A ∈ C([0, T ]; H)), and the equation in (6.167) is satisfied in V 0 for almost all t ∈ ]0, T [. If in addition F ∈ L2 (0, T ; H), then A ∈ C([τ, T ]; V) ∩ L2 (τ, T ; D) , At ∈ L2 (τ, T ; H) ,

(6.177)

286

6

Examples

for all τ ∈ ]0, T [. If A0 ∈ V, we can take τ = 0 in (6.177). Furthermore, if Ft ∈ L2 (0, T ; V 0 ), then At ∈ C([τ, T ]; H) ∩ L2 (τ, T ; V) , A ∈ C([τ, T ]; D) ,

(6.178)

with the choice τ = 0 admissible if A0 ∈ D (so that A1 := curl ζ (curl A0 )+F(0, ·) ∈ H). PROOF A proof of this theorem, based on the Galerkin method, can be given following the same procedure of the proof of theorem 3.9 of chapter 3, where we considered the semilinear parabolic problem. Here, we limit ourselves to establish the necessary a priori estimates that we shall also need in the sequel for the existence of the absorbing balls for the semiflow generated by (6.167). At first, multiplying the equation in (6.167) in H by 2A we obtain d kAk2 + 2hζ (curl A), curl Ai = 2hF, curl AiV 0 ×V ; dt recalling (6.174), we obtain then that d kAk2 + γkAk2V ≤ dt

2 1 2γ kFkV 0

+ 21 γkAk2V ,

(6.179)

from which, integrating in [0, T [, we obtain that A ∈ L∞ (0, T ; H) ∩ L2 (0, T ; V). Hence, At ∈ L2 (0, T ; V 0 ), so that A ∈ C([0, T ]; H). Next, we multiply the equation in (6.167) in H by (et − 1)At , obtaining d (et − 1)N(curl A) dt ≤ 21 (et − 1)kFk2 + 21 (et − 1)kAt k2 + (et − 1)N(curl A) + N(curl A) .

(et − 1)kAt k2 +

Integrating this, and recalling (6.176), we obtain Z t

(eθ − 1)kAt k2 dθ + (et − 1)N(curl A)

0

≤

Z t

0

0

+

Z t (eθ − 1)kFk2 dθ + L kAk2V dθ

Z t

(eθ − 1)N(curl A) dθ .

(6.180)

0

Recalling that A ∈ L2 (0, T ; V), we can then deduce, by means of Gronwall’s inequality, that A ∈ C([τ, T ]; V) and At ∈ L2 (τ, T ; H), τ ∈ ]0, T [. Thus, A ∈ L2 (τ, T ; D), and (6.177) holds. Note that if A0 ∈ V, we do not need to multiply by the factor (et − 1), so that we can take τ = 0 in (6.180). In fact, for future reference we note that, in this case, we have the estimate Z t 0

kAt k2 dθ + γkA(t)k2V ≤ LkA0 k2V +

Z T 0

kFk2 dθ .

(6.181)

6.4

287

Maxwell’s Equations

Finally, we differentiate the equation in (6.167) with respect to t, and multiply the resulting equation in H by 2(et − 1)At , obtaining d (et − 1)kAt k2 + 2(et − 1)hζ 0 (curl A) curl At , curl At i dt = 2(et − 1)hFt , At i + et kAt k2 . From this, recalling (6.173), we obtain d (et − 1)kAt k2 + 2γ(et − 1)kAt k2V ≤ (et − 1)kFt k2 + (2et − 1)kAt k2 , dt and, integrating, (et − 1)kAt k2 + 2γ

Z t 0

≤

Z T

(eθ − 1)kAt k2V dθ

(et − 1)kFt k2 dθ +

Z t

(2eθ − 1)kAt k2 dθ .

(6.182)

0

0

Recalling that At ∈ L2 (0, T ; H) if A0 ∈ V, we conclude that (6.178) holds. 2. If F ∈ L2 (0, +∞; V 0 ), theorem 6.31 shows that problem (6.167) generates a solution operator S = (S(t))t ≥0 , with S(t) : H → H for all t ≥ 0, and S(t) : V → V if F ∈ L2 (0, +∞; H). We now show that, if F is independent of t, these solution operators are actually semiflows, both in H and in V. To this end, it is sufficient to estimate the difference α(t) := A(t) − A(t) = S(t)A0 − S(t)A0 , in H and V. At first, we immediately have that d kαk2 + 2hζ (curl A) − ζ (curl A), curl αi = 0 , dt from which, recalling (6.173), d kαk2 + γkαk2V ≤ 0 . dt

(6.183)

Integrating this inequality we deduce that S(t) is Lipschitz continuous in H, for all t ≥ 0. We now prove that each operator S(t) is also continuous in V; that is, that for all t > 0, A0 ∈ V, and η > 0, there is δ > 0 such that kS(t)A0 − S(t)A0 kV = kA(t) − A(t)kV ≤ η

if kA0 − A0 kV ≤ δ .

Proceeding by contradiction, we assume that there are t0 > 0, A0 ∈ V and η0 > 0 such that for all δ > 0, there is A0 ∈ V with kA0 − A0 kV ≤ δ

but kA(t0 ) − A(t0 )kV ≥ η0 .

(6.184)

288

6

Examples

From (6.183) and (6.169) we deduce that γkα(t0 )k2V ≤ 2kαt (t0 )k kα(t0 )k ≤ 2kαt (t0 )k kα(0)k ≤ 2λF kαt (t0 )k kα(0)kV . (6.185) Consider now any A0 ∈ V, with kA0 − A0 kV ≤ 1. From (6.182) (with Ft ≡ 0) and (6.181) we deduce that kAt (t0 )k2 ≤

2et0 − 1 et0 − 1

Z t0

kAt k2 dθ ≤

0

2et0 − 1 LkA0 k2V + t0 kFk2 ; t 0 e −1

similarly, kAt (t0 )k2 ≤

2et0 − 1 2L(kA0 k2V + 1) + t0 kFk2 . t e 0 −1

Therefore, we deduce that kαt (t0 )k2 ≤

6(2et0 − 1) L(1 + kA0 k2V ) + t0 kFk2 . t e 0 −1

(6.186)

Inserting this into (6.185) we obtain kα(t0 )k2V ≤ C0 kA0 − A0 kV ,

(6.187)

where the constant C0 depends on t0 and kA0 kV , as per (6.186). Let then δ := min{ 21 η02C0−1 , 1} , and A0 ∈ V be determined as in (6.184). Then, kA0 − A0 kV ≤ 1; thus, (6.187) holds, and this implies the contradiction η02 ≤ kα(t0 )k2V ≤ C0 δ ≤ 21 η02 . This allows us to conclude that each solution operator S(t) is also continuous in V. Hence, S is a semiflow in both H and V.

6.4.3 Absorbing Sets and Attractors 1. We now proceed to show the existence of bounded, positively invariant absorbing sets for S, both in H and in V. PROPOSITION 6.32 Assume that F is independent of t. If F ∈ V 0 , there exists a ball B0 ⊂ H, which is positively invariant and absorbing for the semiflow S in H. If F ∈ H, there exists a ball B1 ⊂ V, which is positively invariant and absorbing for the semiflow S in V. PROOF From (6.179) we first obtain, recalling (6.169), the exponential inequality d γ 1 kAk2 + kAk2 ≤ kFk2V 0 , dt 2λF 2γ

6.4

Maxwell’s Equations

289

from which we deduce the existence of a bounded, positively invariant absorbing set B0 for S in H in the usual way. In particular, for any bounded set G ⊂ H, there is M > 0 such that for all A0 ∈ G and all t ≥ 0, kS(t)A0 k =: kA(t)k ≤ M .

(6.188)

Next, we multiply the equation in (6.167) in H by At , to obtain kAt k2 +

d N(curl A) ≤ 12 kFk2 + 12 kAt k2 . dt

(6.189)

From (6.179) we also obtain, recalling (6.176) 2γ 2 1 L N(curl A) ≤ 2γ kFkV 0

+ 2kAk kAt k ≤

2 2 2 1 C 2γ kFk + 2kAk + 2 kAt k .

(6.190)

Summing (6.189) and (6.190), and recalling (6.188), we obtain the exponential inequality d 2 2 1 C N(curl A) ≤ N(curl A) + 2γ 1 + L 2 γ kFk + 2M =: M1 . dt

(6.191)

Because of (6.176), we can deduce from (6.191) the existence of a bounded, positively invariant absorbing set B1 for S in V. 2. Since the injections D ,→ V ,→ H are compact, (6.178) implies that S is asymptotically uniformly compact, both with respect to the topologies of V and H; hence, theorem 2.46 implies that the semiflow S defined by problem (6.167) admits the sets A0 = ω(B0 ) and A1 = ω(B1 ) as global attractors, respectively in H and V. REMARK 6.33 The importance of showing the existence of an attractor for S in V, and not just in H, resides in the fact that we can then deduce the existence of an attractor, in H, for the semiflow S˜ defined by the original first order system (QS), that is, recalling (6.164), by ˜ S(t)B 0 = B(t, ·) := curl A(t, ·) = curl(S(t)A0 ) , where B0 ∈ H, and A0 ∈ V is determined from B0 as the solution of problem (6.164). As for E, note that, by (6.165), E(t, ·) := −At (t, ·) − ∇ϕ, with ∇ϕ ∈ L2 (Ω ) (recall that in the autonomous case, the source term G is independent of t), and At (t, ·) ∈ H for all t > 0, as stated in (6.178). Hence, E(t, ·) remains in a bounded set of L2 (Ω ), for all t ≥ 0. We conclude our presentation of the quasi-stationary Maxwell’s equations with the remark that, since problem (6.167) is quasilinear, the techniques we have developed in chapters 4 and 5 are not directly applicable. Hence, the existence of an exponential attractor for S, even only in H, not to mention that of an inertial manifold, is an open problem.

Chapter 7 A Nonexistence Result for Inertial Manifolds

In this chapter we present a result, due to Mora and Solà-Morales, [MSM87], concerning the nonexistence of inertial manifolds for the semiflow generated by a semilinear dissipative wave equation of the form (3.4).

7.1 The Initial-Boundary Value Problem 1. We consider again the hyperbolic perturbation, for ε > 0, of the Chafee-Infante equation of (5.149) in one space dimension, with f = 0, i.e. εutt + ut − uxx = k(u − u3 ) =: g(u) ,

(7.1)

with k > 0 and t > 0, x ∈ ]0, π[. We supplement (7.1) with the initial conditions u(0, x) = u0 (x) , ut (0, x) = u1 (x) ,

x ∈ ]0, π[ ,

(7.2)

but, in contrast to problem (5.166), we impose homogeneous boundary conditions of Neumann type, i.e. ux (t, 0) = ux (t, π) = 0 ,

t > 0.

(7.3)

We refer to the IBVP (7.1)+(7.2)+(7.3) as problem

(HN ) .

In section 5.8.5 of chapter 5 we have seen that if ε is sufficiently small, the semiflow generated by the corresponding IBVP with Dirichlet boundary conditions admits an inertial manifold. In contrast, Mora and Solà-Morales have shown in [MSM87] that, if instead ε is sufficiently large, the semiflow generated by problem (HN ) cannot admit an inertial manifold of class C1 , containing the global attractor of this semiflow. 2. In order to conform√to Mora and Solà-Morales’ presentation, we introduce again the time rescaling t 7→ εt, which transforms (7.1) into the equation utt + 2αut − uxx = k(u − u3 ) ,

1 α := √ . 2 ε

(7.4)

291

292

7

A Nonexistence Result for Inertial Manifolds

Exactly as in section 3.4, we can prove that problem (HN ) generates a flow in the product space X := H1 (0, π) × L2 (0, π). In fact, arguing as in section 3.4.1 we have the following result. THEOREM 7.1 ˜ t ∈R on X , which admits a global Problem (HN ) generates a continuous flow S˜ = (S(t)) attractor A˜ in X . Moreover, for each compact interval [a, b] ⊂ R, the map ˜ X 3 U0 7→ S(·)U 0 ∈ C([a, b]; X )

(7.5)

is of class C2 and bounded. The proof of this theorem can be established along the same lines as that of theorem 3.20, with the exception of the C2 dependence of the orbits from their initial values, expressed in (7.5). This part can be proven as in Henry, [Hen81, thm. 3.4.4], even if the “space” operator A of the first order formulation of equation (7.4) is not sectorial (see section A.3). Here, we limit ourselves to prove the following forward uniqueness and continuous dependence result in X , which we will need in the sequel. PROPOSITION 7.2 Problem (HN ) is well posed in X , on any compact interval [T0 , T ] ⊂ R (see definition 1.5). PROOF Without loss of generality, we can take T0 = 0. Let Z := C(R; H1 (0, π)) ∩ C1 (R; L2 (0, π)) , and assume that u, u¯ ∈ Z are two solutions of the initial boundary value problem (7.4)+(7.3). Then, their difference z solves the equation ztt + 2αzt − zxx = k(z − u3 + u¯3 ) ,

(7.6)

with homogeneous initial and boundary conditions. Multiplying (7.6) in L2 (0, π) by 2zt , we obtain, as in section 3.4.1: d (kzt k2 + kzx k2 ) + 4αkzt k2 = 2k((1 + (u2 + uu¯ + u¯2 ))z, zt ) dt ≤ 2k(1 + |u2 + uu¯ + u¯2 |∞ )kzk kzt k ,

(7.7)

where k · k denotes, as usual, the norm in L2 (0, π). Since H1 (0, π) ,→ L∞ (0, π), and u, u¯ ∈ Z, given T > 0 there is C0 > 0, depending on u and u, ¯ such that |u2 + uu¯ + u¯2 |∞ ≤ C(kuk21 + kuk ¯ 21 ) =:≤ C0 . Thus, integrating (7.7) in [0,t], 0 < t < T , we obtain ϕ(t) := kzt (t, ·)k2 + kzx (t, ·)k2 + 2α

Z t 0

kzt (θ , ·)k2 dθ

7.2

293

Overview of the Argument

≤ kzt (0, ·)k2 + kzx (0, ·)k2 +C1

Z t

kz(θ , ·)k2 dθ ,

(7.8)

0

with C1 depending on C0 and α. Estimating kz(θ , ·)k2 ≤ 2kz(0, ·)k2 + 2

Z

2

θ

kzt (τ, ·)k dτ

0

≤ 2kz(0, ·)k2 + 2θ

Z θ

kzt (τ, ·)k2 dτ ,

(7.9)

0

we obtain from (7.8) that ϕ(t) ≤ kzt (0, ·)k2 + kzx (0, ·)k2 + 2C1 kz(0, ·)k2t + Cα1

Z t

θ ϕ(θ ) dθ . 0

By Gronwall’s inequality, we deduce then that ϕ(t) ≤ (kzt (0, ·)k2 + kzx (0, ·)k2 + 2Ckz(0, ·)k2t)eC1 t

2

/(2α)

˜ . =: ϕ(t)

(7.10)

Replacing this into (7.9), we have ˜ . kz(t, ·)k2 ≤ 2kz(0, ·)k2 + α1 t ϕ(t) Together with (7.10), this implies the conclusion of proposition 7.2. In particular, solutions of problem (HN ) in X are uniquely determined by their initial values.

7.2 Overview of the Argument Mora and Solà-Morales’ nonexistence result is proven by contradiction; in its general lines, it runs as follows. The second order equation (7.4) is transformed into the first order system in X Ut + αU = AU + G(U) , where A is the linear operator, unbounded on X , formally defined by ! 0 1 , A := α 2 + ∂xx 0

(7.11)

(7.12)

and, for U = (u, u) ˜ ∈ X , G(U) := (0, k(u − u3 )) (compare to (5.114). We define S = (S(t))t ∈R to be the flow generated by (7.11) in X . This flow has, in the phase space X , three stationary states P−1 , P0 and P1 , which are the constant functions P−1 (x) ≡ (−1, 0) , P0 (x) ≡ (0, 0) , P1 (x) ≡ (1, 0) ;

(7.13)

294

7

A Nonexistence Result for Inertial Manifolds

the stationary states P0 and P1 are joined by a heteroclinic orbit γ, which is the graph of a C1 function. By theorem 7.1, translated to the equivalent first order system (7.11), the flow S admits a global attractor A in X ; thus, γ ⊆ A. In particular, denoting by FS the class of the finite dimensional manifolds, which are locally invariant under the flow S, are of class C1 , contain P1 , and are differentiable at P1 , then, for any neighborhood Vη of P1 in X we have that γ ∩ Vη ∈ FS . Suppose now that S admitted a closed inertial manifold M ⊆ X . Then, γ ⊆ A ⊆ M; therefore, M ∩ Vη ∈ FS . The central point of Mora and Solà-Morales’ argument is to show that this cannot happen. They achieve this, by proving that the set FS is at most countable, and, therefore, it contains “too few” elements for M ∩ Vη to be in FS , for a suitable choice of Vη sufficiently small. To show that FS is countable, Mora and Solà-Morales’ first prove that the same is true for the analogous set FR , corresponding to the linear flow R, obtained by linearizing (7.11) at P1 (see (A.5)). Then, they resort to an equivalence result, analogous to the Hartman-Grobman theorem (see theorem A.8), to “transfer” the result back to the original flow S. Linearizing (7.11) at P1 , we obtain the system Ut + αU = AkU , where Ak is the unbounded linear operator in X , formally defined by ! 0 1 . Ak := α 2 − 2k + ∂xx 0

(7.14)

(7.15)

The linear system (7.14) generates a flow in X , which we denote by R = (R(t))t ∈R . Acting as in (5.117), we see that if α is so small, i.e. ε is so large, that α 2 < 2k ,

(7.16)

then Ak admits a sequence of eigenvalues, which are all purely imaginary. In section 7.4 we show that, as a consequence, the set FR consists of at most one countable family; that is, FR = (MRm )m∈N , each manifold MRm being finite dimensional, locally invariant under R, of class C1 , containing P1 , and differentiable at P1 . By a C1 linearization equivalence result, shown in section 7.10, the same holds in a neighborhood Vη of P1 in X , for the original nonlinear problem (7.11). That is, also the set FS consists of at most one countable family FS = (Mm )m∈N . As we discussed above, from this it follows that, if S admitted a closed inertial manifold M, then there must exist m ∈ N such that M ∩ Vη = Mm ∩ Vη . Mora and Solà-Morales’ proceed then to show, by means of a perturbation argument, that this cannot happen. To describe this step of their argument, in section 7.6 we show that system (7.11) can be perturbed in such a way that the corresponding flow S f still admits P−1 , P0 and P1 as stationary states, with P0 and P1 joined by a heteroclinic orbit γ f , contained in the global attractor A f of S f . This perturbation can be made so, that if Vη is a sufficiently small neighborhood of P1 in X , then the

7.3

295

The Linearized Problem

restriction of S f to Vη coincides with that of S. Thus, A f ∩ Vη = A ∩ Vη and, consequently, γ f ∩ Vη = γ ∩ Vη ∈ FS . That is, γ f must converge to P1 along one of the countable manifolds Mm of the family FS . In the final part of their argument, Mora and Solà-Morales’ show then explicitly that this cannot happen; that is, γ f does not converge to P1 along any of the manifolds Mm . From this contradiction, they can then deduce that the flow generated by the original system (7.11) cannot admit an inertial manifold.

7.3 The Linearized Problem 1. We consider the linearized problem (7.14) in X = H1 (0, π) × L2 (0, π). The domain of the operator Ak defined by (7.15) is the space Y := HN2 (0, π) × H1 (0, π) ,

(7.17)

where HN2 (0, π) := {u ∈ H2 (0, π) : ux (0) = ux (π) = 0} . System (7.14) is equivalent to the second order equation utt + 2αut + Lk u = 0 , where Lk is the linear operator on L2 (0, π), with domain H2 (0, π), defined by u 7→ Lk u := −uxx + 2ku . The eigenvalues of Lk , and the corresponding eigenfunctions, are λm = m2 + 2k , wm (x) = cos(mx) ,

m ∈ N,

(7.18)

and the sequence (wm )m∈N forms an orthogonal system in both L2 (0, π) and H1 (0, π). This means that, if for m ∈ N we set Hm := span{wm } , Xm := Hm × Hm ,

(7.19)

then the orthogonal decompositions L2 (0, π) =

∞ M

m=0

Hm , X =

∞ M

Xm

(7.20)

m=0

hold, the first being invariant with respect to Lk . Finally, we recall that, since the domain of Lk , i.e. HN2 (0, π), is compactly imbedded into L2 (0, π), the operator Lk has compact resolvent (see theorem A.24).

296

7

A Nonexistence Result for Inertial Manifolds

2. We now turn our attention to the operator Ak defined by (7.15). Our goal is to choose in the product space X a norm, with respect to which Ak generates a group of unitary operators (see section A.2). More precisely, we assume that (7.16) holds, that is, α 2 < 2k, set α1 := 2k − α 2 , and endow X with the equivalent scalar product defined, for U = (u, u) ˜ and V = (v, v) ˜ ∈ X , by ¯˜ . hU,V iX := hux , v¯x i + α1 hu, vi ¯ + hu, ˜ vi

(7.21)

In (7.21), h · , · i denotes, as usual, the scalar product in L2 (0, π), and the bar denotes complex conjugation (recall that all the eigenfunctions of Ak are complex valued). We claim: PROPOSITION 7.3 Let Y be defined as in (7.17). For all U ∈ Y, RehAkU,UiX = 0. Consequently, the operator iAk (where i is the imaginary unit) is self-adjoint. PROOF 1. Let U = (u, u) ˜ ∈ Y. Recalling (7.21), we compute ¯˜ hAkU,UiX = hu˜x , u¯x i + α1 hu, ˜ ui ¯ + h−α1 u + uxx , ui ¯ ¯ = (hu˜x , u¯x i − hux , u˜x i) − α1 (hu, ui ˜ − hu, ˜ ui) ¯ = (hux , u¯˜x i − hux , u¯˜x i) − α1 (hu, ˜ ui ¯ − hu, ˜ ui) ¯ . Since Re(z − z¯) = 0 for all complex numbers z, we conclude that RehAkU,UiX = 0 as claimed. 2. Analogously, for U = (u, u) ˜ and V = (v, v) ˜ ∈ Y, we compute ¯˜ hiAkU,V iX = hiu˜x , v¯x i + α1 hiu, ˜ vi ¯ + hi(−α1 u + uxx ), vi = −hu˜x , ivx i − α1 hu, ˜ ivi + α1 hu, ivi ˜ + hux , iv˜x i , ˜ + hu, ˜ i(−α1 v + vxx )i hU, iAkV iX = hux , iv˜x i + α1 hu, ivi = hux , iv˜x i + α1 hu, ivi ˜ − α1 hu, ˜ ivi − hu˜x , ivx i . Consequently, hiAkU,V iX = hU, iAkV iX , as claimed. 3. By Stone’s theorem (see theorem A.50), Ak generates a C0 group Z = (Z(t))t ∈R of unitary operators in X . In particular, each Z(t) is an isometry; that is, for all U0 ∈ X and all t ∈ R, kZ(t)U0 kX = kU0 kX ,

(7.22)

where now k · kX denotes the norm in X corresponding to the scalar product (7.21), i.e. kUk2X := kux k2 + α1 kuk2 + kuk ˜ 2

7.3

297

The Linearized Problem

(recall that α1 = 2k − α 2 > 0, by (7.16)). We claim then: PROPOSITION 7.4 For all t ∈ R, R(t) = e−αt Z(t). PROOF Since Z is generated by Ak , for all U0 ∈ X the function t 7→ V (t) := Z(t)U0 solves the initial value problem V 0 = AkV , V (0) = U0 . Then, the function t 7→ U(t) := e−αt V (t) solves the initial value problem U 0 = e−αt (V 0 − αV ) = e−αt (AkV − αV ) = AkU − αU , U(0) = V (0) = U0 . Thus, (7.14) holds and, therefore, U(t) = R(t)U0 . It follows that R(t)U0 = e−αt V (t) = e−αt Z(t)U0 ; since U0 is arbitrary in X , the conclusion follows. PROPOSITION 7.5 Let t ∈ R and m ∈ N. The restriction of Z(t) to the finite dimensional subspace Xm defined in (7.19) is the rotation matrix ! cos(ωmt) ωm−1 sin(ωmt) , Zm (t) := −ωm sin(ωmt) cos(ωmt) where ωm :=

p p m2 + α1 = m2 + 2k − α 2 .

PROOF Recalling (7.19) and (7.18), a generic element U0 ∈ Xm has the form U0 (x) = (α cos(mx), β cos(mx)) ,

α, β ∈ R .

Then, the components of the function t → 7 U(t) := Z(t)U0 := (u(t), u(t)) ˜ solve the initial value problem ( ( u(x, 0) = α cos(mx) ut = u˜ , . u(x, ˜ 0) = β cos(mx) u˜t = −α1 u + uxx Equivalently, u solves the second order initial value problem    utt − uxx + α1 u = 0 u(x, 0) = α cos(mx)   ut (x, 0) = β cos(mx) .

298

7

A Nonexistence Result for Inertial Manifolds

By separation of variables, we easily obtain that ( u(x,t) = (α cos(ωmt) + ωm−1 β sin(ωmt)) cos(mx) , u(x,t) ˜ = (−αωm sin(ωmt) + β cos(ωmt)) cos(mx) .

(7.23)

Since this can be written as U(t) = Zm (t)U0 , the conclusion follows. 4. Setting now am :=

1 (αωm − iβ ) , bm := am , 2ωm

cm := iωm am , dm := −iωm bm ,

we can write (7.23) in the complex form ( u(x,t) = (am eiωm t + bm e−iωm t ) cos(mx) , u(x,t) ˜ = (cm eiωm t + dm e−iωm t ) cos(mx) .

(7.24)

Consequently, if U0 ∈ X , by (7.24) and the second of (7.20) we have the series expansions  ∞   u(x,t) = ∑ (am eiωm t + bm e−iωm t ) cos(mx) , m=0 (7.25) ∞   u(x,t) ˜ = ∑ (cm eiωm t + dm e−iωm t ) cos(mx) . m=0

Since ωm ∈ R for each m ∈ N, it follows from (7.25) and theorem A.98 that for all U0 ∈ X the function R 3 t 7→ U(·,t) = Z(t)U0 ∈ X is almost periodic in t. Thus, by Bochner’s theorem A.97, the set of translations {U(· + τ) : τ ∈ R} is relatively compact in Cb (R; X ), with respect to the sup norm.

7.4 Inertial Manifolds for the Linearized Problem 1. We are now in a position to prove the basic result of this section, which will allow us to show that the flow R = (R(t))t ∈R generated by the linearized problem (7.14) has only at most countably many finite dimensional, locally invariant submanifolds containing the origin, and differentiable there. We start with DEFINITION 7.6 Let M be a global, trivial C1 manifold in a Banach space X ; that is, M is the graph of a function ϕ ∈ C1 (X1 ; X ), with X1 a finite dimensional subspace of X (see definition 5.2). Let x0 = ϕ(ξ0 ) ∈ M, ξ0 ∈ X1 . Consider the affine map Lx0 : X1 → X , defined by Lx0 (ξ ) := x0 + ϕ 0 (ξ0 )(ξ − ξ0 ) ,

ξ ∈ X1 .

7.4

Inertial Manifolds for the Linearized Problem

299

The image of Lx0 is called the TANGENT SPACE to M at x0 , and denoted Tx0 M, if lim

ξ →ξ0

kϕ(ξ ) − Lx0 (ξ )kX = 0. kξ − ξ0 kX

Example 7.7 The set M := {(x, y) ∈ R2 : y = arctan x} is a one dimensional, trivial Lipschitz submanifold of R2 , whose tangent space at x0 = 41 is the set T1/4 M := {(ξ , η) ∈ R2 : 16ξ − 17η + 17 arctan 41 − 4 = 0} .

We have then the following result. PROPOSITION 7.8 Let M ⊆ X be a submanifold of X , such that 0 ∈ M, M is differentiable at 0, and M is invariant with respect to the flow R. Let TM denote the tangent space to M at 0. Then M ⊆ TM. PROOF 1. Let P : X → TM be the projection onto TM, and Q = IX − P. Since TM is tangent to M at 0, lim

U ∈M U→0

kQUkX = 0. kPUkX

(7.26)

Let U0 ∈ M, and V (t) := R(t)U0 . By proposition 7.4 and (7.22), kV (t)kX = e−αt kZ(t)U0 kX = e−αt kU0 kX → 0 as t → +∞. Since V (t) ∈ M for all t ≥ 0, by (7.26) lim

t →+∞

kQV (t)kX = 0. kPV (t)kX

(7.27)

Let U(t) := Z(t)U0 . Again by proposition 7.4, (7.27) implies lim

t →+∞

kQU(t)kX kQV (t)kX = lim = 0. t →+∞ kPV (t)kX kPU(t)kX

Since P and Q are orthogonal, and Z(t) is an isomorphism, we have that 0 ≤ kU0 k2 − kPU(t)k2X = kQU(t)k2X =

kQU(t)k2X kPU(t)k2X kPU(t)k2X

(7.28)

300

7 ≤

A Nonexistence Result for Inertial Manifolds

kQU(t)k2X kQU(t)k2X kU(t)k2X = kU0 k2X . 2 kPU(t)kX kPU(t)k2X

Therefore, by (7.28), lim kPU(t)kX = kU0 kX ,

lim kQU(t)kX = 0 ,

t →+∞

t →+∞

from which lim QU(t) = 0 ,

t →+∞

lim PU(t) = U0 .

t →+∞

(7.29)

Let (τm )m∈N ⊂ R>0 be such that τm → +∞. As we have recalled at the end of the previous section, the sequence (U(· + τm ))m∈N is relatively compact, by Bochner’s theorem; hence, we can extract a subsequence (U(· + τmk ))k∈N , converging uniformly to a functionW ∈ Cb (R; X ). Since P is continuous, we deduce that PU(τmk ) → PW (0) as k → ∞. But then, the second of (7.29) implies that U0 = PW (0) ∈ TM. Thus, M ⊆ TM, as claimed. As an immediate consequence, we have that if M is an invariant submanifold of X , differentiable at 0, then M is in fact a closed, linear subspace of X . Indeed, proposition 7.8 implies that M ⊆ TM; on the other hand, M and TM are homeomorphic, via the mapping M 3 u 7→ ϕ 0 (0)ϕ −1 u ∈ TM, where ϕ is the map whose graph is M, as per definition 5.2 (keep in mind that since ϕ is differentiable at 0, and ϕ −1 is Lipschitz continuous, ϕ 0 (0) is invertible). Hence, M = TM, and the tangent space TM is linear. 2. We now give a further characterization of the linear subspaces of X which are invariant with respect to the flow R. THEOREM 7.9 Assume F ⊆ X is a closed linear subspace, which is invariant with respect to R. There exists a subset N1 ⊆ N such that F=

M

Xm ,

(7.30)

m∈N1

where Xm is as in (7.19). PROOF 1. Since R(t) = e−αt Z(t) and F is linear, F is also invariant with respect to Z. We now show that also the orthogonal complement F ⊥ of F in X is invariant with respect to Z. Indeed, given U0 ∈ F ⊥ , take any V0 ∈ F. Since Z(t) is unitary for all t ∈ R, its adjoint satisfies Z(t)∗ = Z(t)−1 = Z(−t) ;

7.4

Inertial Manifolds for the Linearized Problem

301

therefore, hZ(t)U0 ,V0 iX = hU0 , Z(t)∗V0 iX = hU0 , Z(−t)V0 iX . Since Z(−t)V0 ∈ F because F is invariant, it follows that hZ(t)U0 ,V0 iX = 0; hence, Z(t)U0 ∈ F ⊥ , and F ⊥ is invariant with respect to Z. 2. We next show that for each m ∈ N, F ∩ Xm 6= {0} ⊥

F ∩ Xm 6= {0}

⇐⇒ ⇐⇒

Xm ⊆ F , ⊥

Xm ⊆ F .

(7.31) (7.32)

Consider in fact (7.31). Since Xm does not reduce to the origin, if Xm ⊆ F then obviously F ∩ Xm = Xm also does not reduce to {0}. Conversely, note that F ∩ Xm is a subspace of both F and Xm , which is invariant with respect to Z. However, proposition (7.5) implies that Xm does not contain any such proper subspace. Since F ∩ Xm does not reduce to the origin, it must be F ∩ Xm = Xm , and thus Xm ⊆ F. This proves (7.31); the proof of (7.32) is analogous. 3. Set now, for t ∈ R, ˜ := 1 (Z(t) + Z(−t)). Z(t) 2 We claim that for all m ∈ N, t ∈ R and U ∈ X , ˜ U ∈ Xm ⇐⇒ Z(t)U = cos(ωmt)U ,

(7.33)

p where ωm = m2 + α1 as in proposition 7.5. Indeed, if U ∈ Xm , the identity at the right side of (7.33) follows from (7.23). Conversely, suppose U ∈ X satisfies the right side of (7.33), but U ∈ / Xm . Then, recalling the second of (7.20), there must be r ∈ N, with r 6= m, such that hU,Ur iX 6= 0 for some Ur ∈ Xr , for otherwise, if hU,Ur i = 0 for all r 6= m, ∞

U=

∑ hU,Ur iX Ur = hU,Um iX Um ∈ Xm .

r=0

Since U 6= 0, U is either in F or in F ⊥ . Since these subspaces are both invariant ˜ with respect to Z, Z(t)U is in the same subspace where U is. Hence, from (7.33) we obtain that ˜ hZ(t)U,U r iX = cos(ωm t)hU,Ur iX 6= 0 .

(7.34)

˜ It is now easy to check that each operator Z(t) is also unitary; therefore, since the ˜ map t 7→ Z(t) is symmetric, ˜ ˜ −1Ur iX = hU, Z(−t)U ˜ ˜ hZ(t)U,U r iX = hU, Z(t) r iX = hU, Z(t)U r iX .

(7.35)

˜ But from the part of (7.33) we have already proven, Z(t)U r = cos(ωr t)Ur , because Ur ∈ Xr . From (7.35) and (7.34) we have then that cos(ωmt)hU,Ur iX = cos(ωr t)hU,Ur iX ; since hU,Ur iX 6= 0, this implies that cos(ωmt) = cos(ωr t). Since t ∈ R is arbitrary, this forces ωm = ωr ; since this is not true, we conclude that U ∈ Xm .

302

7

A Nonexistence Result for Inertial Manifolds

4. We now show that for all m ∈ N, either Xm ⊆ F, or Xm ⊆ F ⊥ . Indeed, given U ∈ Xm \ {0}, we can uniquely decompose U = UF ⊕UF⊥ , UF ∈ F , UF⊥ ∈ F ⊥ . Then, by (7.33), for all t ∈ R ˜ Z(t)U = cos(ωmt)U = cos(ωmt)UF + cos(ωmt)UF⊥ . Since also ⊥ ˜ ˜ ˜ Z(t)U = Z(t)U F + Z(t)U F , ⊥ ⊥ ˜ ˜ and Z(t)U F ∈ F, Z(t)U F ∈ F , it follows that ⊥ ⊥ ˜ ˜ Z(t)U F = cos(ωm t)UF , Z(t)U F = cos(ωm t)UF .

(7.36)

Since (7.33) is an equivalence, (7.36) implies that UF and UF⊥ are both in Xm . Hence, UF ∈ F ∩ Xm , and UF⊥ ∈ F ⊥ ∩ Xm . Since at least one of UF , UF⊥ is not zero, the conclusion follows from (7.31) and (7.32). 5. We are now ready to conclude the proof of theorem 7.9. Given F as above, set N1 := {m ∈ N : Xm ⊆ F} . We claim that, then, (7.30) holds. Indeed, it is clear that M

Xm ⊆ F .

(7.37)

m∈N1

To show the inverse inclusion, let U ∈ F. Recalling the second of (7.20), we decompose U=

∑

m∈N1

αmUm +

∑

αmUm =: U 0 +U 00 .

m∈ / N1

By (7.37), U 0 ∈ F; consequently, U 00 = 0, and U = U 0 ∈

L

Xm . Thus, (7.30)

m∈N1

follows, and the proof of theorem 7.9 is complete. 3. As a consequence of theorem 7.9, we deduce THEOREM 7.10 If α 2 < 2k, the flow R = (R(t))t ∈R generated by the linearized equation (7.14) admits at most a countable family of closed, finite dimensional, invariant manifolds which contain the stationary point 0 and are differentiable at this point. PROOF By proposition 7.8, any such manifold M is a closed, linear subspace of X . By theorem 7.9, M, being finite dimensional, is the direct sum of at most a finite number of subspaces of Xm . Since the family of finite subsets of N is countable, the conclusion follows.

7.5 C1 Linearization Equivalence

303

7.5 C1 Linearization Equivalence Theorem 7.10 characterizes a certain class of finite dimensional, invariant manifolds of the flow R generated by the linearized problem (7.14), namely those which contain the stationary point U = 0, and are differentiable at this point. Our goal is now to extend this result to the flow S generated by the original nonlinear problem (7.11), of which (7.14) is the linearization at its stationary state P1 = (1, 0) (regarded as a constant function in X ). We achieve this by generalizing to the infinite dimensional case (i.e. to general Banach spaces) the Hartman-Grobman theorem A.8 on the topological equivalence at the origin of the systems of ODEs x˙ = f (x) ,

x˙ = f 0 (0)x ,

where f ∈ C1 (Rn ; Rn ) and f (0) = 0. We start with the following C1 linearization result. THEOREM 7.11 Let X be a Banach space, and F : X → X be a map, not necessarily linear. Let x0 ∈ X be a fixed point of F, and assume that F is of class C1 in a neighborhood U of x0 . Let L := F 0 (x0 ) (the Fréchet derivative of F at x0 ), and assume that L has a bounded inverse. Assume further that kF 0 (x) − LkL(X ,X ) = O(kx − x0 kX ) kL

−1

as x → x0 (x ∈ U) ,

2

kL(X ,X ) kLkL(X ,X ) < 1 .

(7.38) (7.39)

There exist then a neighborhood V ⊆ U of x0 , with F(V) ⊆ V, and a C1 diffeomorphism Φ : V → Φ(V) ⊆ X , such that Φ(x0 ) = 0, Φ 0 (x0 ) = IX , Φ 0 (x) − IX = O(kx − x0 kX )

as x → x0 (x ∈ V) ,

and ΦF = LΦ

on V .

Moreover, Φ is unique, in the sense that if W ⊆ U and Ψ : W → X also satisfy the same conditions as V and Φ, then Φ ≡ Ψ on V ∩ W. We postpone the proof of this theorem to section 7.10.1, and proceed instead to extend it to a general flow S on X , and its linearization R at a stationary point x0 of S. THEOREM 7.12 Let (S(t))t ∈R be a flow on a Banach space X , such that for each t ∈ R, S(t) is a C2 diffeomorphism on a neighborhood U of a common fixed point x0 (i.e., S(t)x0 = x0

304

7

A Nonexistence Result for Inertial Manifolds

for all t ∈ R). For each t ∈ R, let R(t) := (S(t))0 (x0 ) (the Fréchet derivative of S(t) at x0 ). Assume there is τ ∈ R such that the operators F := S(τ) and L := R(τ) satisfy conditions (7.38) and (7.39) of theorem 7.11. There exist then a neighborhood V ⊆ U of x0 , and a C1 diffeomorphism Φ : V → Φ(V) ⊆ X , such that Φ(x0 ) = 0, Φ 0 (x0 ) = IX , kΦ 0 (x) − IX kL(X ,X ) = O(kx − x0 kX )

as x → x0 (x ∈ V) ,

and the identity ΦS(t) = R(t)Φ holds for all t ∈ R in a ball B(x0 , r) ⊆ V. This ball can be chosen independently of t if t is bounded from below. We prove this theorem in section 7.10.2. We now show that this theorem can be applied to the flows S and R generated, respectively, by problem (7.11) and its linearization (7.14) at the stationary state P1 . Indeed, choosing any τ > 0, we immediately check that F = S(τ) and L = R(τ) do satisfy the assumptions of theorem 7.12. In fact, (7.38) holds because S(τ) is of class C2 , as we know from theorem 7.1; as for (7.39), recall from proposition 7.4 that R(τ) = e−ατ Z(τ), and that Z(τ) is an isometry. Thus, kLkL(X ,X ) = kR(τ)kL(X ,X ) = e−ατ kZ(τ)kL(X ,X ) = e−ατ , kL−1 kL(X ,X ) = kR(−τ)kL(X ,X ) = eατ kZ(−τ)kL(X ,X ) = eατ . It follows that kL−1 kL(X ,X ) kLk2L(X ,X ) = e−ατ < 1 as required. We summarize this conclusion in THEOREM 7.13 If α 2 < k, the flow S generated by problem (7.11) is, near its stationary state P1 , C1 equivalent to its linearization R, i.e. to the flow generated by the linearized problem (7.14). Consequently, there is a neighborhood V of P1 in X , with the property that there is one at most countable family F of submanifolds of X , whose intersection with V are closed, finite dimensional manifolds, each of which is invariant with respect to S, contains the intersection A ∩ V (where A is the global attractor of S in X ), and is differentiable at P1 .

7.6 Perturbations of the Nonlinear Flow 1. In this section we perturb the flow S generated by the nonlinear problem (7.11) in such a way that, on one hand, the long-time behavior of the perturbed flow is the

7.6

Perturbations of the Nonlinear Flow

305

same as that of S in a neighborhood of its stationary state P1 , and, on the other, near P1 the global attractor of the perturbed flow, which must coincide with the global attractor A of S, is not contained in any of the at most countably many manifolds containing P1 . From this we deduce that S cannot admit an inertial manifold M, because otherwise A ⊆ M and, near P1 , M must be one of the at most countably many manifolds containing P1 . 2. Going back to the second order IBVP (HN ), we realize that since the nonlinearity g is independent of x, and the boundary conditions are of Neumann type, the flow S admits a 2-dimensional invariant linear subspace in X , consisting of functions that are independent of the space variable x. More precisely, let L2c (0, π) := {u ∈ L2 (0, π) : u(x) ≡ const. a.e. in ]0, π[} . Then obviously L2c (0, π) ⊆ H1 (0, π), and Xc := L2c (0, π) × L2c (0, π) ⊆ X . Note that if P−1 , P0 and P1 are as in (7.13), then P−1 , P0 and P1 ∈ Xc . We claim: PROPOSITION 7.14 Xc is a 2-dimensional linear subspace of X , invariant with respect to S. PROOF Xc is obviously a linear subspace of X . Let (u0 , u1 ) ∈ Xc , and set (u(t), u(t)) ˜ := S(t)(u0 , u1 ) . Then, u solves (HN ). On the other hand, since u0 and u1 are constant functions, which we can identify with two numbers, we can consider the solution y of the Cauchy problem (

y00 + 2αy0 = k(y − y3 ) y(0) = u0 , y0 (0) = u1 .

(7.40)

Set v(x,t) := y(t). Then, v solves (HN ) with the same initial data; therefore, by uniqueness (proposition 7.2), v = u and u˜ = vt . But (v(·,t), vt (·,t)) ∈ Xc for all t ∈ R; thus, (u(t), u(t)) ˜ := S(t)(u0 , u1 ) ∈ Xc . This means that S(t)Xc ⊆ Xc for all t ∈ R. Conversely, given (u0 , u1 ) ∈ Xc and t¯ ∈ R, let y be the solution of the Cauchy problem (7.40), with the initial values replaced by y(t¯) = u0 , y0 (t¯) = u1 . Then (y(0), y0 (0)) ∈ Xc , and (u0 , u1 ) = S(t¯)(y(0), y0 (0)) ∈ S(t¯)Xc . This means that Xc ⊆ S(t¯)Xc . Thus, Xc is invariant with respect to S. Finally, Xc is obviously 2dimensional, since it is isomorphic to R2 .

306

7

A Nonexistence Result for Inertial Manifolds

3. Since the equation of (7.40) has exactly the same structure of Duffing’s equation (1.52), with λ = 0, we know from section 2.3.3 that there is a heteroclinic orbit γ joining the stationary states P0 and P1 . As in figure 2.3, γ has the shape shown in figure 7.1; note that condition (7.16), i.e. α 2 < 2k, guarantees that P1 is an hyperbolic, asymptotically stable stationary state. v p(u) =

κ 3 2α (u − u )

max v γ max u P0

a ϕ0

u P1 b

Figure 7.1: Heteroclinic orbit joining P0 to P1

Since γ is an orbit of the flow generated by (7.40) (and of S), there are functions t 7→ ϕ(t) and t 7→ ψ(t) such that γ = {(ϕ(t), ψ(t)) : t ∈ R} . In particular, there is t0 ∈ R such that ϕ 0 (t) > 0 , ψ 0 (t) = ϕ 00 (t) > 0

(7.41)

ϕ(t) ≥ ϕ(t0 ) =: ϕ0 ,

(7.42)

for t < t0 , while for t ≥ t0 ,

with 0 < ϕ0 < 1. We fix then two numbers a and b, with 0 < a < ϕ0 < b < 1 ,

(7.43)

and introduce the set of functions J (a, b) := { f ∈ C2 ([0, π] × R) : f (x, u) ≡ 0

if u ≤ a or u ≥ b} .

7.7

Asymptotic Properties of the Perturbed Flow

307

This set is a Banach space with respect to the usual C2 norm. For f ∈ J (a, b), we consider the “perturbed” IBVP    utt + 2αut − uxx = g(u) + f (x, u) (7.44) u(0, x) = u0 (x) , ut (0, x) = u1 (x)   ux (t, 0) = ux (t, π) = 0 , which we transform as usual into its first order formulation Ut + αU = AU + G(U) + F(x,U) ,

(7.45)

with A and G as in (7.11), and F(x,U) := (0, f (x, u)), U ∈ X . This problem, which is autonomous, generates a flow on X ; more precisely, as in theorem 7.1 we have THEOREM 7.15 For each f ∈ J (a, b), problem (7.45) generates a continuous flow S f = (S f (t))t ∈R on X , which admits a compact, global attractor A f in X . Moreover, for each compact interval [t0 ,t1 ] ⊂ R, the map X × J (a, b) 3 (U0 , f ) 7→ S f (·)U0 ∈ C([t0 ,t1 ]; X )

(7.46)

is of class C2 .

7.7 Asymptotic Properties of the Perturbed Flow 1. Theorem 7.13 shows that there is a ball B(P1 , r) in X on which the flow S is C1 -equivalent to its linearization R. If η ≤ η0 =: min{r, 1 − b} , then B(P1 , η) ∩ (]a, b[ ×{0}) = ∅ in X ;

(7.47)

consequently, if f ∈ J (a, b) and U = (u, u) ˜ ∈ B(P1 , η), f (x, u) = 0 for all x ∈ [0, π]. This implies that the restrictions of the flows S f and S to Vη := B(P1 , η) coincide; therefore, the asymptotic behavior of these restrictions are the same. In particular, each S f admits A ∩ Vη as a global attractor in X , and, if S admitted an inertial manifold M, M ∩ Vη is an inertial manifold (with boundary) for S f . Moreover, since P1 is a stable stationary state of both S and S f , there is η 0 ≤ η such that for all t ≥ 0, S(t)B(P1 , η 0 ) = S f (t)B(P1 , η 0 ) ⊆ Vη .

(7.48)

308

7

A Nonexistence Result for Inertial Manifolds

Since η ≤ r, it follows that also the restriction of the flow S f to Vη is C1 -equivalent to the restriction of the linearized flow R to Vη . Hence, as in theorem 7.13, there are only at most countably many submanifolds of X , whose intersection with Vη are closed, finite dimensional manifolds, which are invariant with respect to S f , contain the intersection A ∩ Vη , and are differentiable at P1 . Since S f is independent of f on Vη , this family is actually independent of f ; in fact, it coincides with the family FS of manifolds of the flow S, described in theorem 7.13. 2. Our strategy is then the following. In proposition 7.17 below, we fix a choice of η ∈ ]0, η0 ]. In turn, this choice of η will determine another parameter δ . We show then that if the perturbation f is “sufficiently small”, as measured by this parameter δ , then the corresponding flow S f admits an heteroclinic orbit γ f joining the stationary states P0 to P1 . We know that γ f ⊆ A (since A is the attractor of S f ); thus, if S admitted an inertial manifold M containing A, then, since M is also an inertial manifold for S f , γ f ⊆ M as well. Since γ f converges to P1 , it must enter Vη ; hence, γ f must converge to P1 along one of the manifolds of FS . More precisely, if S admitted an inertial manifold M, then there must exist a manifold Mm ∈ FS , such that γ f ∩ Vη ⊆ Mm . We proceed then to show that there is at least one f ∈ J (a, b) such that this does not happen; i.e., that there is at least one point on the corresponding set γ f ∩ Vη , which is not on any one of the manifolds of FS . This produces a contradiction, which shows that the flow S cannot admit a locally invariant inertial manifold M, containing the attractor A. 3. We now show that when f is small, S f admits an heteroclinic orbit γ f joining P0 to P1 . The choice of a in (7.43) uniquely determines a point U∗ = (a, a) ˜ ∈ γ. For f ∈ J (a, b) and t ∈ R, we let U f (t) := S f U∗ =: (u f (t), u˜ f (t))

(7.49)

(thus, u˜ f = αu f + (u f )t ), and denote by γ f ⊆ X the image of the curve t 7→ U f (t). Finally, for ρ > 0, we denote by Jρ the ball of J (a, b) with center 0 and radius ρ (in the C2 norm). We claim: PROPOSITION 7.16 For all η ∈ ]0, η0 ], there is δ1 ∈ ]0, η] with the property that if δ ∈ ]0, δ1 ] and f ∈ Jδ , then γ f is an heteroclinic orbit for S f , joining its stationary states P0 and P1 . PROOF Since f (·, u) = 0 for u ≤ a, and ϕ is increasing for t ≤ t0 (recall (7.42)), it follows that S f (t)U∗ = S(t)U∗ for all t ≤ t0 . Consequently, lim U f (t) = (0, 0) = P0 .

t →−∞

(7.50)

Since P1 is an asymptotically stable stationary state of S, there is Tη > 0 such that for all t ≥ Tη , S(t)U∗ ∈ B(P1 , η/2) .

(7.51)

7.8

The Nonexistence Result

309

By the continuity of the map (U∗ , f ) 7→ S f (·)U∗ (see (7.46)), (7.51) implies that there is δ1 > 0 such that if f ∈ Jδ1 and Tη ≤ t ≤ Tη + δ1 , S f (t)U∗ ∈ Vη . From (7.48) we deduce then that S f (t)U∗ = S(t)U∗ for all t ≥ Tη . Thus, lim U f (t) = lim S(t)U∗ = P1 .

t →+∞

t →+∞

Together with (7.50), this implies that γ f is an heteroclinic orbit for S f , joining P0 to P1 .

7.8 The Nonexistence Result For η ≤ η0 , let δ ≤ δ1 ≤ η, Tη be as in the proof of proposition 7.16, and Vη as in (7.47). Recall, from our previous discussion, that if S admits an inertial manifold M containing the global attractor A, then there is Mm ∈ FS such that Vη ∩ M = B(P1 , η) ∩ Mm . We now define a map Φ : Jδ → B(P1 , η) ⊆ X by Φ( f ) := S f (Tη )U∗ .

(7.52)

Then, Φ( f ) belongs to Vη ∩ γ f ⊆ Vη ∩ A; hence, Φ( f ) ∈ Vη ∩ Mm ,

(7.53)

for some Mm ∈ FS . Our goal is then to show that if δ is sufficiently small, there are many functions f ∈ Jδ such that (7.53) does not hold, i.e. Φ( f ) cannot belong to any of the at most countably many finite dimensional manifolds Mm of FS , which contain P1 and are differentiable at P1 . To this end, we prepare PROPOSITION 7.17 For η ∈]0, η0 ], let δ1 ∈]0, η] be as in proposition 7.16, and determine Tη , so that (7.51) holds. There exist numbers η and δ , with 0 < δ ≤ η ≤ η0 , and, correspondingly, a dense subset Λδ of Jδ , such that for all f ∈ Λδ , the range of Φ 0 ( f ) is an infinite dimensional subspace of X (recall that Φ 0 ( f ) ∈ L(Jδ ; X )). We postpone the proof of this result to the next section. Proposition 7.17 determines our choice of η; in turn, this choice determines the values of δ (via δ1 of proposition 7.16), and Tη , as well as the neighborhood Vη of P1 . These choices allow us to finally arrive at Mora and Solà-Morales’ nonexistence result:

310

7

A Nonexistence Result for Inertial Manifolds

THEOREM 7.18 Let FS be the at most countable family of closed, finite dimensional C1 manifolds of S, containing the stationary state P1 , and differentiable at P1 . Let δ be as in proposition 7.17, and Φ( f ) be as in (7.52). There is a dense subset Σ of Jδ , with the property that if f ∈ Σ , there is no manifold M ∈ FS containing Φ( f ). PROOF 1. Let FS = (Mm )m∈N be the family of manifolds for S, described in theorem 7.13. For each m ∈ N, define Σm := { f ∈ Jδ : Φ( f ) ∈ / Mm } . We claim that each Σm is nonempty, open, and dense in Jδ . As a consequence of Baire’s category theorem A.19, the set Σ :=

\

Σm

m∈N

would then also be dense in Jδ , and the very definition of Σm would imply that Σ has the desired properties. 2. If Σm were empty, then Φ( f ) ∈ Mm for all f ∈ Jδ , so that the range of Φ 0 ( f ) would be contained in the tangent space to Mm at Φ( f ). Since Mm is finite dimensional, this tangent space is also finite dimensional; thus, if we take f ∈ Λδ , we reach a contradiction with the conclusion of proposition 7.17. A similar argument shows that Σm is dense in Jδ . For, otherwise, there would be a g˜ ∈ Jδ and ε0 > 0 such that kg − gk ˜ Jδ ≥ ε0 for all g ∈ Σm . But this would imply that if g ∈ Jδ and kg − gk ˜ Jδ < ε0 , then g ∈ / Σm . Taking in particular g ∈ Λδ (which is possible since Λδ is dense in Jδ ), we reach again a contradiction with proposition 7.17. 3. Finally, we show that Σm is open, as a consequence of the continuity of the map f 7→ S f . Indeed, let f˜ ∈ Σm . Then, recalling (7.52), Φ( f˜) = S f˜(T )U∗ ∈ / Mm . Since Mm is closed, there is an open neighborhood W of Φ( f˜) in X such that W ⊆ Vη and W ∩ Mm = ∅. Then, Φ( f ) ∈ W for all f sufficiently close to f˜ and, for these f , Φ( f ) ∈ / Mm . Hence, f ∈ Σm , and Σm is open. This concludes the proof of theorem 7.18, under the assumption that proposition 7.17 holds.

7.9 Proof of Proposition 7.17 1. We first compute explicitly the Fréchet derivative Φ 0 ( f ) in L(J (a, b); X ), for f ∈ J (a, b). Thus, let h ∈ J (a, b). Recalling the definition (7.52) of Φ( f ), we compute 1 1 Φ 0 ( f )h = lim (Φ( f + rh) − Φ( f )) = lim (S f +rh (Tη )U∗ − S f (Tη )U∗ ) . (7.54) r→0 r r→0 r

7.9

Proof of Proposition 7.17

311

Now, if u f +rh and u f are as in (7.49), they are, respectively, the solutions of the equations utt + 2αut − uxx = g(u) + f (·, u) + rh(·, u) , vtt + 2αvt − vxx = g(v) + f (·, v) , with common initial values (u∗ , u˜∗ ), and Neumann boundary conditions. Thus, writing u f +rh = u f + rw, we deduce that w satisfies the equation r(wtt + 2αwt − wxx ) = g(u f + rw) − g(u f ) + f (x, u f + rw) − f (x, u f ) + rh(x, u f + rw) . Dividing by r and letting r → 0, we obtain that w solves the linear, nonhomogeneous IBVP  0 0   wtt + 2αwt − wxx = g (u f )w + f (x, u f ) + h(x, u f ) (7.55) w(0, x) = 0 , wt (0, x) = 0   wx (t, 0) = wx (t, π) = 0 . Thus, recalling (7.54), we conclude that Φ 0 ( f )h = (w(Tη , ·), wt (Tη , ·) + αw(Tη , ·)) . 2.

(7.56)

We can now proceed to prove proposition 7.17. We set Λ :=

[

J (a0 , b0 ) ,

(7.57)

a
and, for δ > 0, Λδ := Λ ∩ Jδ . We easily see that Λ is dense in J (a, b); as a consequence, Λδ is dense in Jδ . We now show that if δ is sufficiently small, Λδ satisfies the requirements of the proposition. Thus, we have to show that if f ∈ Λδ , with δ sufficiently small, the range of Φ 0 ( f ) is an infinite dimensional subspace of X . We will achieve this by constructing a sequence (hm )m∈N of linearly independent functions in J (a, b), such that the corresponding images Φ 0 ( f )hm , defined as in (7.56), form a sequence of linearly independent functions in X . To this end, we first adjust the parametrization t 7→ U(t) = (ϕ(t), ψ(t)) of the heteroclinic orbit γ, so that ϕ(0) = a , ψ(0) = a˜ .

(7.58)

Then, (7.41) and (7.42) remain valid on an interval ] − ∞,t0 [, with t0 > 0, since ϕ0 = ϕ(t0 ) > a. In fact, ϕ is strictly increasing on a larger interval ] − ∞, T0 [, where T0 > t0 is the first value of t such that ϕ(t) = 1. In particular, ϕ is increasing on [0, Tη ], where Tη is as in the proof of proposition 7.16. Since ψ = ϕ 0 , we also have that ϕ is convex on ] − ∞,t0 [. Next, we note that if f ∈ Λδ , then f ∈ Λ , so by (7.57) there are numbers a f , b f , such that a < a f < b f < b, and f ∈ J (a f , b f ).

312 3.

7

A Nonexistence Result for Inertial Manifolds

At the end of this section, we shall prove

LEMMA 7.19 Let η0 be as in the beginning of section 7.7, and set m f := min{a f , ϕ0 }. There are η ≤ η0 and, correspondingly, δ ≤ δ1 (recall that δ1 depends on η), with the property that for all f ∈ Λδ , there are numbers c f ∈ ]a, m f ] and t f ∈ ]0,t0 ] such that ϕ(t f ) ≥ c f and ∀ (t, x) ∈ ] − ∞,t f ] × [0, π] : ∀ (t, x) ∈ [t f , +∞[ × [0, π] :

u f (t, x) ≤ ϕ(t) ≤ a f , u f (t, x) ≥ c f .

(7.59) (7.60)

Assuming this lemma to hold, we fix f ∈ Λδ , and proceed with the construction of the desired sequence (hm )m∈N . In addition to the properties described above, we further require that the IBV problems (7.55) corresponding to this sequence be as “simple” as possible; ideally, we would like to be able to solve these problems by simple separation of variables. To this end, we first determine the numbers a f , b f , c f and t f corresponding to f as per lemma 7.19, and remark that since f (x, u) ≡ 0 if u ≤ a f , by the uniqueness of the solutions to problems (7.44) and (7.11) it follows that u f (t, x) = ϕ(t)

(7.61)

for all t ≤ t0 and x ∈ [0, π]. Thus, if we restrict our attention to the interval [0,t0 ], and only consider functions supported in the interval [a, a f ], the equation in (7.55) simplifies into wtt + 2αwt − wxx − g0 (ϕ(t))w = h(x, ϕ(t)) .

(7.62)

As we have stated, we would like to solve this equation by separation of variables. Thus, it is natural to look for a sequence (wm )m∈N of solutions of the form wm (t, x) = µm (t) cos(mx) .

(7.63)

It turns out that we can indeed do so, if we assume each function hm to have the form hm (x, u) = χm (u) cos(mx) ,

(7.64)

with χm of class C2 , and supported in the interval [a, a f ]. With these choices, a direct replacement of (7.64) into (7.62) shows that for each m ∈ N, the function µm in (7.63) should be determined as the solutions to the linear, nonhomogeneous Cauchy problem on [0,t0 ] ( µ 00 + 2α µ 0 + m2 µ − g0 (ϕ(t))µ = χm (ϕ(t)) (7.65) µ(0) = 0 , µ 0 (0) = 0 . Moreover, to ensure that the corresponding functions x 7→ wm (t, x) are linearly independent on [0, π] for all t ≥ 0 (in particular, for t = Tη , as desired), we require that there be at least one t¯ ∈ [0,t0 ] such that µm (t¯) 6= 0 for all m ∈ N (recall (7.63)).

7.9

313

Proof of Proposition 7.17

4. It remains now to construct the functions (χm )m∈N . To this end, we first define, for m ∈ N, a function t 7→ ζm (t) on [0,t0 ] as the solution of the homogeneous Cauchy problem ( ζ 00 + 2αζ 0 + m2 ζ − g0 (ϕ(t))ζ = 0 (7.66) ζ (t f ) = 0 , ζ 0 (t f ) = 0 . Then, recalling that, as we have previously remarked, the function t 7→ ϕ(t) is invertible on ] − ∞,t0 ], and that t f ≤ t0 , for m ∈ N we set χm (u) := (σ 00 ζm + 2σ 0 ζm0 + 2ασ 0 ζm ) , (7.67) t=ϕ −1 (u)

where σ : R → R is a C2 function such that σ (0) ≡ 0 for t ≤ 0 and σ (t) ≡ 1 for t ≥ ϕ −1 (c f ). With these choices, each function χm is indeed supported in [a, c f ], because if u ≤ a (respectively, if u ≥ c f ), then t = ϕ −1 (u) ≤ ϕ −1 (a) = 0 (respectively, t = ϕ −1 (u) ≥ ϕ −1 (c f )), and, therefore, σ (t) = 0 (resp., σ (t) = 1). In either case, σ 0 (t) = σ 00 (t) = 0, and χm (u) = 0. We can now verify that, for each m ∈ N, the function µm (t) := σ (t)ζm (t) does solve the Cauchy problem (7.65), with χm defined by (7.67). Indeed, since σ (0) = σ 0 (0) = 0, the initial conditions µm (0) = µm0 (0) = 0 are taken. Next, recalling (7.66), we compute that the left side of the equation in (7.65) equals σ 00 (t)ζm (t) + 2σ 0 (t)ζm0 (t) + 2ασ 0 (t)ζm (t) ,

(7.68)

while its right side equals, by (7.67), χm (ϕ(t)) = (σ ζm + 2σ ζm + 2ασ ζm ) 00

0 0

0

,

t=ϕ −1 (ϕ(t))

which is exactly (7.68). Finally, we note that µm (t f ) = σ (t f )ζm (t f ) = 1 . In fact, ϕ(t f ) ≥ c f by lemma 7.19; therefore, t f ≥ ϕ −1 (c f ), and σ (t f ) = 1. 5. In conclusion, proposition 7.17 is completely proven, as soon as we see that lemma 7.19 holds. To prove this lemma, we first show that for each η ∈ ]0, η0 ] and f ∈ Λ (the set defined in (7.57)), there is a constant Kη > 0 such that for all t ∈ [0, Tη ], max |u f (t, x) − ϕ(t)| ≤ Kη F(1 − e−t ) ,

0≤x≤π

(7.69)

where F := k f kC2 ([0,π]) , and Tη is determined so that (7.51) holds. Note that K depends on η, at least via Tη . Because of the imbedding H1 (0, π) ,→ C([0, π]), to obtain (7.69) it is sufficient to estimate the norm of the difference z(t, ·) := u f (t, ·) −

314

7

A Nonexistence Result for Inertial Manifolds

ϕ(t) in H1 (0, π). To this end, we resort to (by now familiar) energy estimates on z, which satisfies the IBVP    ztt + 2αzt − zxx = g(u f ) − g(ϕ) + f (x, u f ) (7.70) z(0, x) = 0 , zt (0, x) = 0   zx (t, 0) = zx (t, π) = 0 . Set E(t) := kzt (t, ·)k2 + αhz(t, ·), zt (t, ·)i + α 2 kz(t, ·)k2 + kzx (t, ·)k2 . Acting as in the proof of proposition 7.2, we easily obtain, after multiplication of the equation of (7.70) in L2 (0, π) by 2zt and αz, and addition of the two resulting identities, that E satisfies on [0, Tη ] the estimate p d E(t) ≤ 2C1 F E(t) + 2C2 E(t) , dt

(7.71)

where C1 is independent of t, but C2 depends on Tη , u f and ϕ, via the quantities max ku f (t, ·)kH1 (0,π) ,

0≤t ≤Tη

max kϕ(t)k .

0≤t ≤Tη

From (7.71) we obtain p dp E(t) ≤ C1 F +C2 E(t) ; dt since E(0) = 0, integration of (7.72) yields p E(t) ≤ CC1 F(eC2 t − 1) ≤ 2

C1 C2 Tη (1 − e−C2 t ) . C2 Fe

(7.72)

(7.73)

By Sobolev’s imbedding and Schwarz’ inequalities, there is C3 > 0 such that p kz(t, ·)kC0 ([0,π]) ≤ C3 E(t) ; (7.74) hence, (7.69) follows from (7.74) and (7.73), with Kη := C1C2−1C3 eC2 Tη . 6. Let now t0 be as in (7.42). Without loss of generality, we can assume that Tη ≥ t0 . Consider first the case 0 ≤ t ≤ t0 . Then, since ϕ is convex on [0,t0 ], for 0 < s < t ≤ t0 we have, recalling (7.58), ϕ(t) ≥ ϕ(s) + ϕ 0 (s)(t − s) ≥ ϕ(s) + ϕ 0 (0)(t − s) = ϕ(s) + a(t ˜ − s) .

(7.75)

Let now f ∈ Λδ , so that F ≤ δ . By (7.69), and (7.75) with s = 21 t we obtain that for all x ∈ [0, π] u f (t, x) ≥ ϕ(t) − Kη F(1 − e−t ) ≥ ϕ(t) − Kη δt ≥ ϕ( 21 t) + 12 at ˜ − Kη δt .

7.9

315

Proof of Proposition 7.17

If a˜ ≥ 2Kη δ , this implies that u f (t, x) ≥ ϕ( 21 t) ,

0 ≤ t ≤ t0 .

Consider next the case that t0 ≤ t ≤ Tη . Keeping in mind that ϕ is still increasing on [0, Tη ], we choose t = t0 and s = 21 t0 in (7.75), and obtain u f (t, x) ≥ ϕ(t) − Kη F(1 − e−t ) ≥ ϕ(t0 ) − Kη δ ≥ ϕ( 21 t0 ) + 21 at ˜ 0 − Kη δ . If also at ˜ 0 ≥ 2Kη δ , this implies that u f (t, x) ≥ ϕ( 21 t0 ).

(7.76)

Since ϕ( 12 t0 ) ≤ ϕ(t0 ) = ϕ0 , if η is sufficiently small (7.76) also holds for all t ≥ Tη , because u f (x,t) ∈ Vη . In fact, it is sufficient that η ≤ min{η0 , 1 − ϕ0 }. In conclusion, we have established that if η ≤ min{R, 1 − b, 1 − ϕ0 }

(7.77)

and δ≤

a˜ min{1,t0 } , 2Kη

(7.78)

then ( u f (t, x) ≥

ϕ( 21 t) ϕ( 12 t0 )

if 0 ≤ t ≤ t0 , if t0 ≤ t .

Thus, to conclude the proof of lemma 7.19 it is sufficient to take ( t0 if a f ≥ ϕ0 , c f := ϕ( 21 t f ). t f := −1 ϕ (a f ) if a f ≤ ϕ0 , Indeed, the inequality t f ≤ t0 holds trivially if a f ≥ ϕ0 , while if a f < ϕ0 , it follows from ϕ(t f ) = a f ≤ ϕ0 = ϕ(t0 ) and the invertibility of ϕ on ] − ∞,t0 ]. The inequality c f ≤ m f follows from ( ϕ(t0 ) = ϕ0 = m f if ϕ0 ≤ a f , c f = ϕ( 21 t f ) ≤ ϕ(t f ) = af = mf if ϕ0 ≥ a f . Since t f ≤ t0 , (7.59) follows from (7.61) and ( ϕ(t0 ) = ϕ0 ≤ a f u f (t, x) = ϕ(t) ≤ ϕ(t f ) = af

if ϕ0 ≤ a f , if ϕ0 ≥ a f .

316

7

A Nonexistence Result for Inertial Manifolds

Finally, (7.60) follows from (7.76) and u f (t, x) ≥ ϕ( 12 t0 ) ≥ ϕ( 21 t f ) = c f . With this, the proof of lemma 7.19 is complete. Note that we first define η by (7.77), and then δ by (7.78); thus, δ depends on η, not only through δ1 of proposition 7.16, but also via Kη , which depends on Tη . In conclusion, proposition 7.17 and, consequently, Mora and Solà-Morales’ nonexistence theorem 7.18 are now completely proven.

7.10 The C1 Linearization Equivalence Theorems In this section we prove theorems 7.11 and 7.12.

7.10.1 Equivalence for a Single Operator 1. We start with the case of a single operator. For convenience, we reproduce the statement of theorem 7.11, and abbreviate k · k := k · kX . THEOREM 7.20 Let X be a Banach space, and F : X → X be a map, not necessarily linear. Let x0 ∈ X be a fixed point of F, and assume that F is of class C1 in a neighborhood U of x0 . Let L := F 0 (x0 ) (the Fréchet derivative of F at x0 ), and assume that L has a bounded inverse. Assume further that kF 0 (x) − LkL(X ;X ) = O(kx − x0 k) kL

−1

as x → x0 (x ∈ U) ,

2

kL(X ;X ) kLkL(X ;X ) < 1 .

(7.79) (7.80)

There exist then a neighborhood V ⊆ U of x0 , with F(V) ⊆ V, and a C1 diffeomorphism Φ : V → Φ(V) ⊆ X , such that Φ(x0 ) = 0, Φ 0 (x0 ) = IX , kΦ 0 (x) − IX kL(X ;X ) = O(kx − x0 k)

as x → x0 (x ∈ V) ,

(7.81)

and ΦF = LΦ

on V .

(7.82)

Moreover, Φ is unique, in the sense that if W ⊆ U and Ψ : W → X also satisfy the same conditions as V and Φ, then Φ ≡ Ψ on V ∩ W. PROOF 1. We first show that condition (7.80) implies that L is a strict contraction, i.e. that kLkL(X ;X ) < 1. Assuming otherwise, for x and y ∈ X let u = L−1 x and

7.10

The C1 Linearization Equivalence Theorems

317

v = L−1 y. Then kx − ykX ≤ kL−1 kL(X ;X ) kLkL(X ;X ) kx − ykX <

1 kx − ykX ≤ kx − ykX , kLkL(X ;X )

which is a contradiction. 2. We want to determine Φ as the solution of the operator equation (7.82), which we rewrite as Φ = L−1 ΦF .

(7.83)

Without loss of generality, we can assume that x0 = 0; thus, F(0) = 0. Near 0 we can then write F = L + G, and Φ = I +Ψ , where I is the identity in X , and (7.83) becomes I +Ψ = L−1 (I +Ψ )(L + G) = L−1 (L + G +Ψ (L + G)) = I + L−1 G + L−1Ψ (L + G) , that is, Ψ = L−1Ψ (L + G) + L−1 G = K1 (Ψ ) + L−1 G =: K(Ψ ) .

(7.84)

In this equation, L and G are known, and we seek to determine the unknown Ψ as a fixed point of the map K defined by (7.84). To this end, for δ > 0 we denote by Bδ the ball of X of center 0 and radius δ , and introduce the space H := {ψ ∈ C1 (B(0, δ ); X ) : ψ(0) = 0, ψ 0 (0) = 0, kψ 0 (x)kL = O(kxk) (x → 0)} . H is easily seen to be a Banach space with respect to the norm kψkH :=

sup 0
kψ 0 (x)kL . kxk

(7.85)

We immediately have PROPOSITION 7.21 Let G := F − L. For all δ > 0, the restrictions of G and L−1 G to B(0, δ ), which we still denote by G and L−1 G, are in H. PROOF To show that G ∈ H, note first that G(0) = F(0) − L(0) = 0, and G0 (0) = F 0 (0) − L = 0, because L = F 0 (x0 ) = F 0 (0). By (7.79), with x0 = 0, kG0 (x)kL(X ;X ) = kF 0 (x) − LkL(X ;X ) = O(kxk) as x → 0. Consequently, (L−1 G)(0) = L−1 (G(0)) = 0, (L−1 G)0 (0) = L−1 (G0 (0)) = 0, and k(L−1 G)0 (x)kL(X ;X ) = kL−1 (G0 (x))kL(X ;X )

318

7

A Nonexistence Result for Inertial Manifolds ≤ kL−1 kL(X ;X ) kG0 (x)kL(X ;X ) = O(kxk)

as x → 0. Thus, L−1 G ∈ H as well. As a consequence of proposition 7.21, the maps K1 and K introduced in (7.84) are well defined, because if δ is sufficiently small, L + G maps B(0, δ ) into itself. In fact, since G ∈ H, kG(x)k = O(kxk) as x → 0; thus, if kxk ≤ δ , kLx + G(x)k ≤ kLkL(X ;X ) kxk + ckxk2 ≤ (kLkL(X ;X ) + cδ )kxk .

(7.86)

Since kLkL(X ;X ) < 1, there is δ0 > 0 such that kLkL(X ;X ) + cδ ≤ 1 for all δ ∈ ]0, δ0 ]. Hence, (7.86) implies that (L + G)(x) ∈ B(0, δ ) if x ∈ B(0, δ ) and δ ≤ δ0 . Consequently, the composition of functions ψ ∈ H with L + G is defined. 3. We now show that if δ is sufficiently small, K is a strict contraction of H into itself. PROPOSITION 7.22 Let δ ≤ δ0 . For all ψ ∈ H, K1 (ψ) ∈ H. If also δ kL−1 kL(X ;X ) ≤ 1, then K is a strict contraction on H. PROOF 1. Since both ψ and G ∈ H, [K1 (ψ)](0) = [L−1 ψ(L + G)](0) = L−1 (ψ(0)) = 0 . Next, we compute the Fréchet derivative of K1 (ψ) at a generic point x ∈ X . Using the symbol ◦ to denote the composition of linear operators, by the chain rule we compute that (K1 (ψ))0 (x) = L−1 ◦ ψ 0 (Lx + G(x)) ◦ (L + G0 (x)) .

(7.87)

Thus, (K1 (ψ))0 (0) = L−1 ◦ ψ 0 (0) ◦ (L + G0 (0)) = 0 . As in the proof of proposition 7.21, kL + G0 (x)kL(X ;X ) ≤ kLkL(X ;X ) + O(kxk) ≤ kLkL(X ;X ) + c1 δ0 ;

(7.88)

since G is continuous and G(0) = 0, Lx + G(x) → 0 as x → 0. Thus, since ψ ∈ H, we obtain from (7.86), (7.87) and (7.88) k(K1 (ψ))0 (x)kL ≤ kL−1 kL(X ;X ) kψ 0 (Lx + G(x))kL(X ;X ) kL + G0 (x)kL(X ;X ) ≤ kL−1 kL(X ;X ) c2 kLx + G(x)kL(X ;X ) (kLkL(X ;X ) + c1 δ0 ) ≤ kL−1 kL(X ;X ) c2 (kLkL(X ;X ) + cδ )kxk(kLkL(X ;X ) + c1 δ0 ) =: c3 kxk .

7.10

The C1 Linearization Equivalence Theorems

319

This completes the proof that K1 (ψ) ∈ H. Since L−1 G ∈ H by proposition 7.21, it follows that also K(ψ) ∈ H. 2. To show that K is a contraction, let ψ1 , ψ2 ∈ H. Then K(ψ1 ) − K(ψ2 ) = K1 (ψ1 ) − K1 (ψ2 ), and, recalling (7.85), kK1 (ψ1 ) − K1 (ψ2 )kH = sup x∈Bδ x6=0

1 k(K1 (ψ1 ))0 (x) − (K1 (ψ2 ))0 (x)kL(X ;X ) . (7.89) kxk

Let x ∈ B(0, δ ). From (7.87) we have (K1 (ψ1 ))0 (x)−(K1 (ψ2 ))0 (x) = L−1 ◦(ψ10 (Lx+G(x))−ψ20 (Lx+G(x)))◦(L+G0 (x)); thus, recalling (7.86) and (7.89), and that Lx + G(x) ∈ B(0, δ ) if x ∈ B(0, δ ), 1 k(K1 (ψ1 ))0 (x) − (K1 (ψ2 ))0 (x)kL(X ;X ) kxk kψ 0 (Lx + G(x)) − ψ20 (Lx + G(x))kL(X ;X ) ≤ kL−1 kL(X ;X ) 1 kLx + G(x)k kLx + G(x)k · kL + G0 (x)kL(X ;X ) kxk

(7.90)

≤ kL−1 kL(X ;X ) kψ10 − ψ20 kH (kLkL(X ;X ) + ckxk)(kLkL(X ;X ) + c1 kxk) ≤ kψ10 − ψ20 kH kL−1 kL(X ;X ) kLk2L(X ;X ) + (c + c1 )kL−1 kL(X ;X ) kLkL(X ;X ) + cc1 kL−1 kL(X ;X ) kxk2 ≤ kψ10 − ψ20 kH (kL−1 kL(X ;X ) kLk2L(X ;X ) + cL kxk) ,

(7.91)

where cL := (c + c1 )kL−1 kL(X ;X ) kLkL(X ;X ) + cc1 kL−1 kδ0 . Because of (7.80), we can find δ1 ∈ ]0, δ0 ] such that kL−1 kL(X ;X ) kLk2L(X ;X ) + cL δ1 = 12 (1 + kL−1 kL(X ;X ) kLk2L(X ;X ) ) =: ρ < 1 . With this choice of δ1 , we conclude from (7.88) and (7.91) that if δ ≤ δ1 , kK1 (ψ1 ) − K1 (ψ2 )kH ≤ ρkψ1 − ψ2 kH . Thus, K is a strict contraction in H, as claimed. 4. From proposition 7.22 it follows that equation (7.84) has a unique solution Ψ ∈ H; setting Φ = I +Ψ , we conclude that (7.83), and therefore (7.82), hold. Set then V := B(0, δ ). Since Ψ ∈ C1 (B(0, δ ); X ), by theorem A.45 we conclude that Φ is a diffeomorphism between V and its image. Consequently, the proof of theorem 7.11 is complete.

320

7

A Nonexistence Result for Inertial Manifolds

7.10.2 Equivalence for Groups of Operators We can now apply theorem 7.11 to extend the C1 -equivalence theorem to a group of operators. For convenience, we reproduce the statement of theorem 7.12: THEOREM 7.23 Let (S(t))t ∈R be a flow on a Banach space X , such that for each t ∈ R, S(t) is a C2 diffeomorphism on a neighborhood U of a common fixed point x0 (i.e., S(t)x0 = x0 for all t ∈ R). For each t ∈ R, let R(t) := (S(t))0 (x0 ) (the Fréchet derivative of S(t) at x0 ). Assume there is τ ∈ R such that the operators F := S(τ) and L := R(τ) satisfy conditions (7.38) and (7.39) of theorem 7.11. There exist then: a neighborhood V ⊆ U of x0 , and a C1 diffeomorphism Φ : V → Φ(V) ⊆ X , such that Φ(x0 ) = 0, Φ 0 (x0 ) = IX , kΦ 0 (x) − IX kL(X ;X ) = O(kx − x0 kX )

as x → x0 (x ∈ V) ,

(7.92)

and the identity ΦS(t) = R(t)Φ

(7.93)

holds for all t ∈ R in a ball B(x0 , r) ⊆ V. This ball can be chosen independently of t if t is bounded from below. PROOF 1. By theorem 7.20, there are a neighborhood V of x0 and a C1 diffeomorphism Φ : V → Φ(V) such that Φ(x0 ) = 0, Φ 0 (x0 ) = I and (7.81), (7.82) hold. V and Φ depend on τ, via the positions F = S(τ) and L = R(τ). Fix t ∈ R, t 6= τ, and set Vt := S(−t)V, Φt := R(−t)ΦS(t). We claim that Φt satisfies, on Vt , the same conditions satisfied by Φ on V. In fact: at first, Φt (x0 ) = R(−t)ΦS(t)x0 = 0 , because S(t)x0 = 0. Next, recalling that R(θ ) is a linear operator for each θ ∈ R, that (S(t))0 (x0 ) = R(t), and that Φ 0 (x0 ) = I, we compute that Φt0 (x0 ) = R(−t) ◦ Φ 0 (S(t)x0 ) ◦ (S(t))0 (x0 ) = R(−t) ◦ Φ 0 (x0 ) ◦ R(t) = R(−t)R(t) = I .

(7.94)

2. To show that Φt satisfies (7.92), we write again Φ = I +Ψ , Ψ ∈ H, as in the proof of theorem 7.11. Then, from (7.94), for all z ∈ Vt : Φt0 (z) − I = R(−t) ◦ (I +Ψ 0 (S(t)z)) ◦ (S(t))0 (z) − I = [R(−t) ◦ (S(t))0 (z) − I] + R(−t) ◦Ψ 0 (S(t)z) ◦ (S(t))0 (z) =: Λ1 (z) + Λ2 (z) . (7.95) Since S(t) ∈ C2 (X ; X ), we can estimate kΛ1 (z)kL(X ;X ) = kR(−t)((S(t))0 (z) − R(t))kL

The C1 Linearization Equivalence Theorems

7.10

321

≤ kR(−t)kL(X ;X ) k(S(t))0 (z) − (S(t))0 (x0 )kL(X ;X ) ≤ c1 (t)kz − x0 k .

(7.96)

As for Λ2 , we note that since z ∈ Vt = S(−t)V, the point x := S(t)z is in V; thus, by (7.92), as in (7.96), kΛ2 (z)kL(X ;X ) = kR(−t)kL(X ;X ) kΨ 0 (x)kL(X ;X ) k(S(t))0 (z)kL(X ;X ) ≤ ckR(−t)kL(X ;X ) kx − x0 k(kz − x0 k + kR(t)kL(X ;X ) ) .

(7.97)

By Taylor’s formula, kx − x0 k = kS(t)z − S(t)x0 k = kR(t)(z − x0 )k + O(kz − x0 k2 ) ; thus, if kz − x0 k ≤ δ , we obtain from (7.97) kΛ2 (z)kL(X ;X ) = ckR(−t)kL(X ;X ) (kR(t)kL(X ;X ) + c2 (t)δ )kz − x0 k ≤ c3 (t)kz − x0 k . Putting this and (7.96) back into (7.95), we deduce that kΦt0 (z) − IkL(X ;X ) ≤ c4 (t)kz − x0 k ,

(7.98)

i.e. that (7.92) holds for Φt in Vt (recall that x0 = S(−t)x0 ∈ Vt ). 3. Finally, we show that (7.93) also holds for Φt . In fact, we prove more, that is, that for all t ∈ R, Φt = R(−t)Φt S(t) = R(−τ)Φt S(τ) .

(7.99)

Indeed, recalling that Φt := R(−t)ΦS(t), L = R(τ) and F = S(τ), from (7.83) we compute R(−τ)Φt S(τ) = R(−τ)R(−t)ΦS(t)S(τ) = R(−t)R(−τ)ΦS(τ)S(t) = R(−t)ΦS(t) = Φt . Thus, by uniqueness, Φt = Φ on V ∩ Vt , and we can take this common operator as the required Ψ ; in particular, (7.93) follows from (7.99). 4. The dependence of c4 on t in (7.98) shows that, in general, identity (7.93) holds on a neighborhood of x0 whose diameter may depend on t. To show that this neighborhood can be determined independently of t if t is bounded from below, fix t0 ∈ R, and consider only t ≥ t0 . It is then sufficient to show that the set W :=

\

S(−t)V

t ≥t0

is a neighborhood of x0 . To this end, we split ! W = W1 ∩ W2 :=

\ t0 ≤t ≤t0 +τ

S(−t)V

! ∩

\ t ≥t0 +τ

S(−t)V ,

322

7

A Nonexistence Result for Inertial Manifolds

and show separately that W1 and W2 are neighborhoods of x0 . Assume for the moment that it is possible to find r > 0 such that for all t ∈ [t0 ,t0 +τ], S(t)B(x0 , r) ⊆ V .

(7.100)

Then B(x0 , r) ⊆ S(−t)V for all t ∈ [t0 ,t0 +τ]; hence, W1 ⊇ B(x0 , r) and, therefore, W1 is a neighborhood of x0 . As for W2 , we first recall that, by construction, S(τ)V ⊆ V. Arguing by induction, it follows that S(mτ)V ⊆ V for all integer m ≥ 1. Given then t ≥ t0 + τ, we write t = t0 + θ + mτ, with m ≥ 1 and 0 ≤ θ < τ, so that, by (7.100), S(t)B(x0 , r) = S(mτ)S(t0 + θ )B(x0 , r) ⊆ S(mτ)V ⊆ V . Consequently, W2 ⊇ B(x0 , r), and W2 is a neighborhood of x0 . 5. To conclude the proof of theorem 7.23 it is therefore sufficient to find r > 0 such that (7.100) holds. To achieve this, let η > 0 be such that B(x0 , η) ⊆ V. Since S(t) is ˜ > 0 such that continuous and S(t)x0 = x0 for all t ∈ R, there is ρ(t) ˜ S(t)B(x0 , ρ(t)) ⊆ B(x0 , η) .

(7.101)

˜ : (7.101) holds} , ρ(t) := sup{ρ(t) r := inf{ρ(t) : t0 ≤ t ≤ t0 + τ} .

(7.102) (7.103)

Set

We claim that r > 0. Assuming this, then for all t ∈ [t0 ,t0 + τ], S(t)B(x0 , r) ⊆ B(x0 , η) ⊆ V , and (7.100) holds. Clearly, r > 0 if ρ were a continuous function, because ρ is positive over the compact interval [t0 ,t0 + τ]. Otherwise, assume that r = 0. By (7.103), there is a minimizing sequence (θm )m∈N , such that ρ(θm ) → 0. Since (θm )m∈N ⊆ [t0 ,t0 +τ], there exists a subsequence, which we still denote by (θm )m∈N , converging to some θ ∈ [t0 ,t0 + τ]. Since the map (x,t) 7→ S(t)x is continuous, and S(θ )x0 = x0 , given η as above there is β > 0 such that kx − x0 k + |t − θ | < β

=⇒

kS(t)x − x0 k < η .

(7.104)

We can choose m0 ∈ N such that kθm − θ k < 21 β and ρ(θm ) ≤ 41 β for all m ≥ m0 . By (7.102), there is ym ∈ S(θm )B(x0 , 21 β ) such that kym − x0 k ≥ η .

(7.105)

Let xm ∈ B(x0 , 21 β ) be such that ym = S(θm )xm . Then, x = xm and t = θm satisfy the left side of (7.104). Consequently, kS(θm )xm − x0 k = kym − x0 k < η , contradicting (7.105). Thus, r > 0, and (7.100) follows. The proof of theorem 7.23 is now complete.

Appendix: Selected Results from Analysis

In this appendix we put together a number of definitions and results on various topics in Analysis, that we have referred to in the previous chapters. We present these results mostly without proof, but always with at least one reference, indicating where a proof can be found. In these citations, a format like “sct. 1.2.3” refers to subsection 3 of section 2 of chapter 1.

A.1 Ordinary Differential Equations In this section we report some well known results on the well-posedness and stability of classical and generalized solutions of ODEs in RN .

A.1.1 Classical Solutions We consider the system of ODEs for a vector valued unknown function R 3 t 7→ x(t) ∈ RN x˙ = f (t, x) ,

(A.1)

with f defined at least on the product I × U, I ⊆ R an interval, and U ⊆ RN an open domain. We also assign to (A.1) the initial condition x(t0 ) = x0 ,

(t0 , x0 ) ∈ I × U .

(A.2)

Sufficient conditions for the local well-posedness of classical solutions to the initial value problem (IVP in short) (A.1)+(A.2) are well known: THEOREM A.1 Let f ∈ C(I × U; RN ). There exists a closed neighborhood {(t, x) ∈ I × U : |t − t0 | ≤ α , kx − x0 k ≤ β } =: I0 × B0 of (t0 , x0 ), and a function x ∈ C1 (I0 ; B0 ), solution of the IVP (A.1)+(A.2). If in addition f satisfies the t-uniform Lipschitz condition in x ∃ L > 0 ∀ (t, x), (t, x) ¯ ∈ I ×U :

k f (t, x) − f (t, x)k ¯ ≤ L kx − xk ¯ ,

323

324

Appendix: Selected Results from Analysis

then the solution x is uniquely determined by the initial values (t0 , x0 ). In fact, x depends continuously on (t0 , x0 ), in the sense that if t 7→ x(t) and t 7→ x(t) ¯ are solutions of (A.1), corresponding to initial values (t0 , x0 ) and (t¯0 , x¯0 ), and defined on neighborhoods I0 and I¯0 of t0 and t¯0 respectively, then there exists C > 0, independent of (t0 , x0 ) and (t¯0 , x¯0 ), such that for all t ∈ I0 ∩ I¯0 , kx(t) − x(t)k ¯ ≤ C (|t0 − t¯0 | + kx0 − x¯0 k) . PROOF See e.g. Coddington-Levinson, [CL55, ch. 1]. By repeated applications of the local existence theorem A.1, local solutions can be extended to a so-called MAXIMAL SOLUTION, that is, a solution of the IVP (A.1)+(A.2) defined on a maximal interval I 0 ⊆ I, with t0 ∈ I 0 . If I 0 = I, we say that the IVP (A.1)+(A.2) can be solved GLOBALLY in I. As the following result shows, extension of a local solution to a global one can be achieved if a time-independent bound on the local solution is available. Such bounds are usually established by means of so-called A PRIORI estimates on the local solutions. THEOREM A.2 Assume there is M > 0 with the property that if I 0 ⊆ I is any interval containing t0 , and x ∈ C1 (I 0 , RN ) is a solution of the IVP (A.1)+(A.2) on I 0 , then sup kx(t)k ≤ M .

(A.3)

t ∈I 0

Then, the IVP (A.1) + (A.2) can be solved globally in I. PROOF See e.g. Coddington-Levinson, [CL55, sct. 2.1]. The bound in estimate (A.3) is called a priori, because M has to be found independently of any particular interval I 0 on which a local solution is determined. As we have seen in section 2.9 of chapter 2, a priori estimates can be established by means of integral or differential inequalities, such as the Gronwall or the exponential inequalities of propositions 2.62 and 2.64.

A.1.2 Generalized Solutions More generally, we can consider GENERALIZED SOLUTIONS to the IVP (A.1) + (A.2), in the sense that we require the ODEs (A.1) to be satisfied only almost everywhere in t. In this context, it is natural to look for solutions of (A.1) that are absolutely continuous. DEFINITION A.3 Let I ⊆ R be an interval, and X = RN (or, more generally, a reflexive Banach space). A function f : I → X is said to be ABSOLUTELY CONTIN -

A.1

Ordinary Differential Equations

325

UOUS if for all ε > 0 there is δ > 0 such that for any finite choice of nonoverlapping subintervals ]ai , bi [⊂ I, i = 1, . . . , r, r

r

∑ |ai − bi | ≤ δ

=⇒

i=1

∑ k f (ai ) − f (bi )kX ≤ ε .

i=1

We denote by AC(I; X ) the space of absolutely continuous functions from I into X . Recalling then from sections A.4 and A.6 below the definitions of the Lebesgue spaces L1 (I) and L1 (I; X ), we have the following THEOREM A.4 Let I ⊆ R be an interval. A function f : I → X is absolutely continuous if and only if f is differentiable almost everywhere in I, with f 0 ∈ L1 (I; X ). PROOF If X = RN , see e.g. Rudin, [Rud74, ch. 8]. For a generalization to functions valued into a reflexive Banach space, see Komura, [Kom70]. We can then state a result, usually known as C ARATHÉODORY’s theorem, which describes sufficient conditions for the existence of generalized solutions to the IVP (A.1)+(A.2). THEOREM A.5 Assume that f (·, x) is measurable in I for each x ∈ U, and that f (t, ·) is continuous on U for each t ∈ I. Assume further that there is a function ϕ ∈ L1 (I), such that for almost all t ∈ I, sup k f (x,t)k ≤ ϕ(t) . x∈U

Then, there exists I 0 ⊆ I, containing t0 , and a function x ∈ AC(I 0 ; U), which is a generalized solution of the IVP (A.1) + (A.2) on I 0 . PROOF See e.g. Coddington-Levinson, [CL55, sct. 2.1].

A.1.3 Stability for Autonomous Systems We now restrict our attention to systems of ODEs that are AUTONOMOUS, i.e. systems (A.1) in which f is independent of t; that is, of the form x˙ = f (x) .

(A.4)

We take t0 = 0 in the initial condition (A.2), and assume that the IVP (A.4)+(A.2) has a global solution on the right interval [0, +∞[. To emphasize the dependence of

326

Appendix: Selected Results from Analysis

this solution on its initial value, we write it as x(·, x0 ). Finally, we call the image of the map t 7→ x(t, x0 ) in RN the ORBIT starting at x0 . DEFINITION A.6 A point x ∈ RN is a STATIONARY POINT (or, an EQUILIBRIUM POINT ), for the autonomous system (A.4) if f (x) = 0. A stationary point x0 of S is said to be: 1. S TABLE, if for any neighborhood U of x0 there is a neighborhood V ⊂ U of x such that any solution t 7→ x(t, x1 ) with x1 ∈ V is such that x(t, x1 ) ∈ U for all t ≥ 0; 2. U NSTABLE, if it is not stable; 3. A SYMPTOTICALLY STABLE, if x0 is stable and there is a neighborhood V of x such that any orbit starting in V converges to x0 , i.e. if for all x1 ∈ V, lim x(t, x1 ) = x0 .

t →+∞

This definition corresponds to the so-called stability in the sense of Lyapunov. The following is a well known criterion for the stability of the stationary points of autonomous systems. THEOREM A.7 Assume that f ∈ C1 , and let x0 be a stationary point of the ODE (A.4). Consider the matrix A := f 0 (x0 ), and assume that all the eigenvalues of A have nonzero real part. Then: 1. If the real part of all the eigenvalues of A are negative, x0 is asymptotically stable; 2. If A has at least one eigenvalue with positive real part, x0 is unstable. PROOF See e.g. Coddington-Levinson, [CL55, sct. 13.1]. The proof of theorem A.7 is based on a result, known as the H ARTMAN -G ROB on the topological equivalence, near the stationary point x0 , of system (A.4) and its linearization at x0 , i.e. the system MAN EQUIVALENCE THEOREM ,

y˙ = A(y − x0 ) ,

A := f 0 (x0 ) ,

(A.5)

of which x0 is also a stationary point. THEOREM A.8 Assume that f ∈ C1 , and that f (x0 ) = 0. Let A := f 0 (x0 ), and assume that all the eigenvalues of A have nonzero real part. There exist then two neighborhoods U and

A.1

Ordinary Differential Equations

327

V of x0 in RN , and a homeomorphism Φ : U → V, such that for all x1 ∈ U, there is a neighborhood I of 0 in R, such that for all t ∈ I, Φ(x(t, x1 )) = x0 + etA (Φ(x1 ) − x0 ) . That is, Φ maps orbits of the “original” system (A.4), which are near x0 , into orbits of the linearized system (A.5), which also are near x0 . PROOF See e.g. Perko, [Per91, sct. 2.8]. We remark that, in the proof of the Hartman-Grobman theorem, it is often assumed that x0 = 0. This is without loss of generality; for, otherwise, we can consider the change of unknown u(t) := x(t) − x0 . Then, u satisfies the ODE u˙ = f (u + x0 ) =: g(u) , and the function t 7→ u(t) ≡ 0 is a solution of this ODE, since g(0) = f (x0 ) = 0; finally, g0 (0) = f 0 (x0 ) = A. In the scalar case, i.e. when N = 1, theorem A.7 just means that if x0 is a stationary point, then x0 is asymptotically stable if f 0 (x0 ) < 0, while it is unstable if f 0 (x0 ) > 0. For example, consider the ODE x˙ = x(1 − x) .

(A.6)

The stationary points are x0 = 0 and x1 = 1, with f 0 (0) = 1 and f 0 (1) = −1. Thus, x0 is unstable, while x1 is asymptotically stable, as it is easy to see by direct analysis of the ODE (A.6). We conclude this section with the celebrated P OINCARÉ -B ENDIXON THEOREM, which gives a complete description of the ω-limits sets of the orbits of the autonomous system (A.4) in the planar case N = 2. THEOREM A.9 Let Ω ⊆ R2 be an open domain, and f ∈ C1 (Ω ; R2 ). Assume that the orbit γ+ = (x(t))t ≥0 of a solution x of (A.4) is contained in a compact subset of Ω , and that its ω-limit set ω(γ+ ) contains no stationary points. Then, either γ+ or ω(γ+ ) is a periodic orbit (called LIMIT CYCLE). PROOF See e.g. Coddington-Levinson, [CL55, ch. 16]. Roughly speaking, theorem A.9 states that either γ+ is already a periodic orbit (in which case ω(γ+ ) = γ+ ), or γ+ converges to ω(γ+ ). Typically, γ+ “spirals” around ω(γ+ ).

328

Appendix: Selected Results from Analysis

A.2 Linear Spaces and their Duals In this section, we assume as known the definitions of Banach and Hilbert spaces, and that of linear maps between linear spaces. Unless otherwise stated, the linear structure of all the spaces we consider refers to the scalar field R, and all Banach spaces we consider are assumed to be separable. If X is a Banach space, we denote by k · kX the norm in X and, if in addition X is a Hilbert space, we denote by h ·, · iX the inner product in X . Most of the material we report can be found in Dautray and Lions, [DL88, ch. VI], or Kato, [Kat95, chs. 3 and 5], to which we refer for additional details and for the proof of those results we occasionally state without giving an explicit justification.

A.2.1 Orthonormal Bases in Hilbert spaces DEFINITION A.10

Let X be a Hilbert space, and e := (em )m∈N a sequence in X .

1. e is an ORTHONORMAL SYSTEM in X if for each j, k ∈ N, ( he j , ek iX =

1 0

if j = k . if j = 6 k

2. e is TOTAL in X , if the only x ∈ X such that hx, e j iX = 0 for all j ∈ N is x = 0. 3. A total orthonormal system in X is called a TOTAL BASIS of X . PROPOSITION A.11 The elements of an orthonormal system are linearly independent. The span of a total system is dense in X , and every nontrivial Hilbert space contains a total basis (e j ) j∈N . Relative to this basis, every x ∈ X admits the F OURIER SERIES EXPANSION ∞

x=

∑ hx, e j iX e j ,

(A.7)

j=0

with the series converging in X . Moreover, the following PARSEVAL IDENTITY holds: ∞

kxk2X =

∑ |hx, e j iX |2 .

j=0

PROOF See e.g. Dautray-Lions, [DL88, sct. VI.1.6.3].

A.2

329

Linear Spaces and their Duals

A.2.2 Dual Spaces and the Hahn-Banach Theorem 1. Dual spaces. DEFINITION A.12 The TOPOLOGICAL DUAL X 0 of X is the set of all linear and continuous functions from X into R. We denote by X 0 × X 3 (x0 , x) 7→ hx0 , xiX 0 ×X ∈ R (or, when the space X is fixed and there is no danger of confusion, simply by hx0 , xi) the so-called DUALITY PAIRING between X and its dual. That is, hx0 , xi is the value of the real number which is the image of x ∈ X under the linear function x0 ∈ X 0 . The dual space X 0 is itself a Banach space, with respect to the norm kx0 kX 0 := sup x∈X x6=0

|hx0 , xiX 0 ×X | = sup |hx0 , xiX 0 ×X | kxkX x∈X

(A.8)

kxkX =1

(see e.g. Yosida, [Yos80, sct. IV.7]). The following result, known as the H AHN -BANACH existence of nontrivial linear functionals.

THEOREM ,

guarantees the

THEOREM A.13 Let X be a normed linear space on R, X0 a subspace of X , and x00 ∈ X00 . There exists x0 ∈ X 0 , such that for all x ∈ X0 , hx0 , xiX 0 ×X = hx00 , xiX 0 ×X0 0

(i.e., x0 is an extension of x00 ), and kx0 kX 0 = kx00 kX 0 . 0

PROOF See e.g. Yosida, [Yos80, sct. IV.5]. 2. Biduals. Since the dual X 0 of a Banach space X is itself a Banach space, we can in turn consider its dual (X 0 )0 , which is called the BIDUAL space of X , and denoted by X 00 . The so-called “ CANONICAL” INJECTION j : X → X 00 is defined by X 3 x 7→ x00 = j(x)

def.

⇐⇒

∀ x0 ∈ X 0 ,

hx00 , x0 iX 00 ×X 0 = hx0 , xiX 0 ×X . (A.9)

The image j(X ) of X into X 00 is a linear subspace, in general proper, of X 00 . DEFINITION A.14 Let X be a Banach space, and X 00 its bidual, defined by (A.9). X is said to be REFLEXIVE if j(X ) = X 00 . In this case, j is an isomorphism. Thus, if X is reflexive, it can be identified to its bidual X 00 by the canonical isomorphism j defined in (A.9).

330

Appendix: Selected Results from Analysis

3. Weak and weak∗ convergence. DEFINITION A.15 Let X be a normed linear space on R, and X 0 its topological dual. 1. A sequence (xm )m∈N ⊂ X is said to x0 ∈ X 0 ,

CONVERGE WEAKLY

hx0 , xm i → hx0 , xi

to x ∈ X if for all

in R .

0 )m∈N ⊂ X 0 is said to CONVERGE WEAKLY∗ to x0 ∈ X 0 if for all 2. A sequence (xm x ∈ X, 0 , xi → hx0 , xi hxm

in R .

THEOREM A.16 Let X be a normed linear space on R, and X 0 its topological dual. Then: 1. Every weakly convergent sequence in X (respectively, every weakly∗ convergent sequence in X 0 ) is bounded. 2. If in addition X is reflexive, every bounded sequence in X (respectively, in X 0 ) contains a weakly (respectively, a weakly∗ ) convergent subsequence. PROOF See e.g. Yosida, [Yos80, scts. V.1, V.2].

A.2.3 Linear Operators in Banach Spaces In this section we review the definitions and most important results concerning various types of bounded and unbounded operators. In particular, we recall the spectral theory of linear operators with compact inverse. Again, we assume that all spaces X , Y, etc. we consider are at least separable Banach spaces. 1. Bounded, unbounded, and closed operators. DEFINITION A.17 Let X and Y be two Banach spaces, and A : X → Y be a linear operator. 1. A is said to be BOUNDED, if there is a constant K such that for all x ∈ X , kAxkY ≤ KkxkX .

(A.10)

Otherwise, A is said to be UNBOUNDED. 2. We denote the set of all linear, bounded operators A : X → Y by L(X , Y).

A.2

Linear Spaces and their Duals

331

3. If Y = X , and A : X → X is a linear operator, but not necessarily bounded, the DOMAIN of A is the set dom(A) := {x ∈ X : Ax ∈ X } . 4. An unbounded operator A : X → Y is said to be CLOSED if whenever (xm )m∈N ⊂ dom(A) is a sequence such that xm → x in X

and

Axm → y in Y

as m → +∞, then x ∈ dom(A) and y = Ax. It is clear from (A.10) that a bounded linear operator is continuous; moreover, L(X , Y) is itself a Banach space, endowed with the norm kAkL(X ,Y ) := sup x∈X x6=0

kAxkY = sup kAxkY . kxkX x∈X

(A.11)

kxkX =1

Likewise, if A : X → X is closed, dom(A) is a Banach space, with respect to the graph norm defined by kuk2dom(A) := kuk2X + kAuk2X ,

u ∈ dom(A) .

(For a proof, see e.g. Dautray-Lions, [DL88, ch. VI].) In particular, recalling definition A.12, we have that X 0 = L(X , R), and (A.8) is in accord with (A.11). PROPOSITION A.18 Assume A ∈ L(X , Y) is bijective. Then A−1 ∈ L(Y, X ). PROOF See e.g. Dautray-Lions, [DL88, sct. VI.1.2]. Finally, we recall that many results on linear operators are based on the following result, known as BAIRE ’ S CATEGORY LEMMA: THEOREM A.19 Let X be a complete metric space, and assume that (An )n∈N is a sequence of nonempty, open subsets of X , each dense in X . Then, the set A :=

∞ \

An

n=0

is also dense in X . PROOF See e.g. Dautray-Lions, [DL88, sct. VI.1].

332

Appendix: Selected Results from Analysis

2. Isometries and unitary operators. DEFINITION A.20 Let X and Y be Hilbert spaces, and A : X → Y be a linear operator. A is said to be: 1. An ISOMETRY, if for all x ∈ X , kAxkY = kxkX . 2. A UNITARY OPERATOR if it is a surjective isometry. Note that an isometry is obviously injective, and the inverse of a unitary operator is unitary. 3. Resolvent, spectrum, eigenvalues and eigenvectors. DEFINITION A.21 Let X be an Hilbert space, and A : X → X be a linear operator, not necessarily bounded. Let I denote the identity in X . Then: 1. The RESOLVENT SET of A is the set ρ(A) := {λ ∈ C : λ I − A is bijective in X } . 2. For λ ∈ ρ(A), the linear operator R(A, λ ) := (λ I − A)−1 : X → X is called the RESOLVENT of A. 3. The POINT SPECTRUM of A is the set σp (A) := {λ ∈ C : λ I − A is not injective in X } . 4. Each λ ∈ σp (A) is called an such that

EIGENVALUE

of A, and each x ∈ dom(A) \ {0}

(λ I − A)x = 0 is called an EIGENVECTOR of A, corresponding to the eigenvalue λ . 4. Compact operators. DEFINITION A.22 Let X and Y be two Banach spaces. A bounded operator A ∈ L(X , Y) is said to be COMPACT, if the image (Axm )m∈N of any sequence (xm )m∈N bounded in X contains a subsequence (Axmk )k∈N , converging in Y. We denote the set of compact operators in L(X , Y) by K(X , Y). It is easy to see that K(X , Y) is a closed subspace of L(X , Y), with respect to the topology of the uniform convergence of operators (i.e., with respect to the norm (A.11)).

A.2

Linear Spaces and their Duals

333

Of particular importance are operators whose inverse are compact. More generally: DEFINITION A.23 Let X be a Hilbert space, and A : X → X be a linear operator, with nonempty resolvent set. A is said to have COMPACT RESOLVENT, if its resolvent R(A, λ ) is compact for all λ ∈ ρ(A). THEOREM A.24 Let X be a Hilbert space, and A : X → X be a linear operator, with nonempty resolvent set. Then, A has compact resolvent if and only if the injection j : dom(A) → X is compact. Moreover, the existence of just one λ ∈ ρ(A) such that R(A, λ ) is compact is sufficient to guarantee that A has compact resolvent. PROOF See e.g. Engel-Nagel, [EN00, sct. 2.4.d].

A.2.4 Adjoint of a Bounded Operator 1. The notion of the ADJOINT of a bounded operator is a straightforward generalization of that of the transpose of a matrix in RN . DEFINITION A.25 Let X and Y be Banach spaces, with duals X 0 and Y 0 , and let A ∈ L(X , Y). The ADJOINT (or TRANSPOSE) of A is the operator A0 : Y 0 → X 0 defined by Y 0 3 y0 7→ x0 = A0 y0

def.

⇐⇒

∀x ∈ X :

hx0 , xiX 0 ×X = hy0 , AxiY 0 ×Y .

(A.12)

It is immediate to see that A0 ∈ L(Y 0 , X 0 ). Moreover, the adjoint of a compact operator is compact (see e.g. Kato, [Kat95, sct. III.4.2]). 2. If X and Y are Hilbert spaces, and A ∈ L(X , Y), we can identify X and Y with their duals X 0 and Y 0 , by means of Riesz’ representation theorem (see e.g. Yosida, [Yos80, sct. III.6]). Then, we can define a linear operator A∗ : Y → X , by the composition A∗ := ρX−1 ◦ A0 ◦ ρY ,

(A.13)

where ρX : X → X 0 and ρY : Y → Y 0 are, respectively, the Riesz’ isomorphisms identifying X and Y with their duals. The operator A∗ defined in (A.13) is also called the ADJOINT of A. Clearly, A∗ ∈ L(Y, X ), and identity (A.12) reads hA∗ y, xiX = hy, AxiY , for all x ∈ X and y ∈ Y. In fact, recalling that ρX is defined by the identity hρX x, yiX 0 ×X = hx, yiX ,

x, y ∈ X ,

(A.14)

334

Appendix: Selected Results from Analysis

and analogously for ρY , recalling (A.12) we have hA∗ y, xiX = hρX−1 A0 ρY y, xiX = hA0 ρY y, xiX 0 ×X = hρY y, AxiY 0 ×Y = hy, AxiY . Thus, when X is a Hilbert space, and Y = X , it makes sense to give the following DEFINITION A.26 Let X be a Hilbert space, and A ∈ L(X , X ). The operator A is said to be SELF - ADJOINT if its adjoint, defined in (A.13), is such that A∗ = A. Recalling (A.14), a self-adjoint operator A ∈ L(X , X ) satisfies the identity ∀ x, y ∈ X :

hAx, yiX = hx, AyiX .

(A.15)

THEOREM A.27 Let X and Y be Hilbert spaces, and A ∈ L(X , Y) be a unitary operator. Then its adjoint A∗ satisfies the equivalent identities A∗ A = IX , AA∗ = IY , A−1 = A∗ , where IX and IY are, respectively, the identity operators in X and Y. PROOF See e.g. Kato, [Kat95, sct. 5.2]. 3. More generally, let X be a reflexive Banach space, Y = X 0 and A ∈ L(X , X 0 ). Then, the transpose operator A0 is in L(X 00 , X 0 ), and we can define an operator A∗ : X → X 0 , by the composition A∗ := A0 ◦ j ,

(A.16)

where j : X → X 00 is the canonical injection of X into its bidual, defined in (A.9). This operator is also called the ADJOINT of A; recalling (A.9), this terminology is justified by the identities hA∗ x, yiX 0 ×X = hA0 j(x), yiX 0 ×X = h j(x), AyiX 00 ×X 0 = hAy, xiX 0 ×X , for all x, y ∈ X . Consequently, we can give DEFINITION A.28 Let X be a reflexive Banach space, and A ∈ L(X , X 0 ). The operator A is said to be SELF - ADJOINT if its adjoint, defined in (A.16), is such that A∗ = A. Recalling (A.14), a self-adjoint operator A ∈ L(X , X 0 ) satisfies the identity hAx, yiX 0 ×X = hAy, xiX 0 ×X ,

∀ x, y ∈ X .

A.2

Linear Spaces and their Duals

335

A.2.5 Adjoint of an Unbounded Operator Let now X be a Hilbert space, and assume that A : X → X is a linear operator, but not necessarily bounded. To define the adjoint A∗ of A, also as an unbounded operator on X , we proceed as follows. First, we introduce the set X0 := {v ∈ X : (∃ w ∈ X ∀ u ∈ dom(A) : hAu, viX = hu, wiX )} ,

(A.17)

and note that if v ∈ X0 , the map dom(A) 3 u 7→ hAu, viX is continuous on dom(A) with respect to the norm induced by X . Indeed, this follows from the estimate |hAu, viX | = |hu, wiX | ≤ kwkX kukX =: Cw kukX . If we assume that dom(A) is dense in X , the element w in (A.17) is uniquely determined by v. This justifies the following DEFINITION A.29 Let X be a Hilbert space, and A a linear operator on X , with domain dom(A) dense in X . Define X0 as in (A.17). The linear operator A∗ : X → X , with domain dom(A∗ ) = X0 , defined by A∗ v := w ,

v ∈ X0 ,

is called the ADJOINT of A in X . Note that, in general, A∗ is an unbounded operator if so is A. DEFINITION A.30 Let X be a Hilbert space, and A a linear operator on X , with domain dom(A) dense in X . A is said to be: 1) SYMMETRIC, if dom(A) ⊆ dom(A∗ ), and for all u, v ∈ dom(A), hAu, viX = hAv, uiX .

(A.18)

(Note that the left side of (A.18) equals hu, A∗ viX .) 2) SELF - ADJOINT if dom(A) = dom(A∗ ), and (A.18) holds for all u, v ∈ dom(A).

A.2.6 Gelfand Triples of Hilbert Spaces 1. We consider two Hilbert spaces V and H, with V ,→ H densely and continuously. In accord with (A.10), the continuity of the injection means that there is a constant C such that for all u ∈ V, kukH ≤ CkukV .

(A.19)

336

Appendix: Selected Results from Analysis

2. Let H0 and V 0 denote, respectively, the topological duals of H and V. The adjoint j∗ of the injection j : V → H is then a natural injection of H0 into V 0 , defined, for h0 ∈ H0 , by the identity h j∗ h0 , uiV 0 ×V := hh0 , juiH0 ×H ,

u∈V.

It follows then that j∗ is injective, and the image j∗ (H0 ) is dense in V 0 . Hence, H0 can be identified to a dense subspace of V 0 . Since H0 can also be identified to H, by means of Riesz’ representation theorem, we finally arrive at the sequence of continuous injections V ,→ H ≡ H0 ,→ V 0 . In this case, we call the three spaces V, H and V 0 a G ELFAND TRIPLE. Finally, we often assume that the injection V ,→ H is not only continuous, but also compact, in the sense of definition (A.23); this means that every sequence in V which is bounded with respect to the norm of V must contain a subsequence, converging in H.

A.2.7 Linear Operators in Gelfand Triples 1. Let now V ,→ H ,→ V 0 be a Gelfand triple, and A ∈ L(V, V 0 ). Since V is dense in H, we can consider A as an unbounded operator in H, with domain dom(A) := {u ∈ V : Au ∈ H} . Then, the identity hAu, viV 0 ×V = hAu, viH

(A.20)

holds for all u ∈ dom(A) and v ∈ V. This implies the following PROPOSITION A.31 Let A be as described. Then, A is self-adjoint as an operator from V into V 0 (i.e., in the sense of definition A.28), if and only if A is self-adjoint as an operator in H (i.e., in the sense of definition A.30 ). PROOF This is a consequence of (A.15) and (A.20), from which it follows that for all u, v ∈ dom(A), hAu, viH = hAu, viV 0 ×V = hAv, uiV 0 ×V = hAv, uiH .

2.

In the same conditions, we also have

A.2

337

Linear Spaces and their Duals

THEOREM A.32 Assume that A is a bijection from V into V 0 . Then A, as an operator in H, is invertible on all of H, and its inverse A−1 is linear and continuous (i.e., A−1 ∈ L(H, H)), and self-adjoint. If in addition the injection V ,→ H is compact, A−1 is a compact operator. PROOF We recall that A−1 is defined, as an operator in L(V 0 , V). Thus, for any h ∈ H ,→ V 0 , A−1 h is defined in V, hence in H. The linearity of A is clear; its continuity follows from the estimate kA−1 hkH ≤ C1 kA−1 hkV ≤ C1 kA−1 kL khkV 0 ≤ C1 kA−1 kLC2 khkH , where C1 and C2 are determined, respectively, by the continuity of the injections V ,→ H and H ,→ V 0 (as in (A.19)). To show that A−1 is self-adjoint, let u, v ∈ H. Then, recalling that A is self-adjoint, hu, A−1 viH = hu, A−1 viV 0 ×V = hAA−1 u, A−1 viV 0 ×V = hA∗ A−1 v, A−1 uiV 0 ×V = hAA−1 v, A−1 uiV 0 ×V = hv, A−1 uiV 0 ×V = hv, A−1 uiH . Finally, let (un )n∈N be a bounded sequence in H. For each n ∈ N, let vn := A−1 un . Since the sequence (un )n∈N is also bounded in V 0 , the sequence (vn )n∈N is bounded in V. Thus, if V ,→ H compactly, there is a subsequence (vnk )k∈N converging in H. Since vnk = A−1 unk , this means that A−1 is compact. 3. We now come to the question of the solvability of an abstract equation of the form Au = f ,

(A.21)

where A ∈ L(V, V 0 ), and f ∈ V 0 . DEFINITION A.33 be: 1.

POSITIVE ,

Let V be a Banach space. An operator A ∈ L(V, V 0 ) is said to

if for all u ∈ V, hAu, uiV 0 ×V ≥ 0 .

2.

STRICTLY POSITIVE ,

if for all u ∈ V \ {0}, hAu, uiV 0 ×V > 0 .

3.

COERCIVE ,

(A.22)

if there is a constant α such that for all u ∈ V, hAu, uiV 0 ×V ≥ αkuk2V .

(A.23)

338

Appendix: Selected Results from Analysis

Note that a coercive operator is strictly positive. Moreover, we have the following result, known as the L AX -M ILGRAM LEMMA: THEOREM A.34 Let V be a Hilbert space, and A ∈ L(V, V 0 ). If A is self-adjoint and coercive, A is an isomorphism between V and V 0 . Consequently, for all f ∈ V 0 problem (A.21) has a unique solution u ∈ V. PROOF See e.g. Dautray-Lions, [DL88, sct. VI.3.2.5]. Note that the map V × V 3 (u, v) 7→ hAu, viV 0 ×V is bilinear and continuous. Thus, we can more generally consider a bilinear continuous map a : V × V → R, and the associated problem of finding, for given f ∈ V 0 , a solution u ∈ V of the problem ∀v ∈ V :

a(u, v) = h f , viV 0 ×V .

(A.24)

This problem, which is called the VARIATIONAL FORMULATION of problem (A.21), can be solved by means of the following version of the Lax-Milgram lemma. THEOREM A.35 Let a be a bilinear continuous map on V × V as above, and assume that a is coercive, in the sense that (compare to (A.23)) there is α > 0 such that for all u ∈ V, a(u, u) ≥ αkuk2V . Then for all f ∈ V 0 , problem (A.24) is uniquely solvable in V. PROOF See e.g. Dautray-Lions, [DL88, sct. VII.1].

A.2.8 Eigenvalues of Compact Operators The following theorem describes the structure of the point spectrum of a compact operator. THEOREM A.36 Let X be a Hilbert space, and L ∈ L(X , X ) be a self-adjoint, strictly positive, compact operator. Then: 1. The point spectrum σp (L) of L is not empty; in fact, either kLkL(X ;X ) ∈ σp (L) or −kLkL(X ;X ) ∈ σp (L). 2. All eigenvalues of L are real and strictly positive, and have finite multiplicity.

A.2

339

Linear Spaces and their Duals

3. These eigenvalues can be ordered into a nonincreasing sequence (µk )k∈N ⊂ R>0 , such that lim µk = 0 .

k→+∞

(A.25)

4. Eigenvectors of L corresponding to distinct eigenvalues are orthogonal. 5. In fact, L admits a complete orthonormal system of eigenvectors. PROOF A proof of part of this theorem can be found e.g. in Zeidler, [Zei95, sct. 4.2], except for the fact that the eigenvalues of L are at most countable and strictly positive. The countability of the eigenvalues is a consequence of the compactness of L; see e.g. Kato, [Kat95, sct. 3.6.7]. The positivity of the eigenvalues is a consequence of the strict positivity of L. Indeed, if u 6= 0 is an eigenvector of L corresponding to the real eigenvalue µ, from (A.22) we deduce that 0 < hLu, uiX = hµu, uiX = µkuk2X .

From theorem A.36 we immediately deduce THEOREM A.37 Let V ,→ H ,→ V 0 be a Gelfand triple of Hilbert spaces, with the injection V ,→ H compact. Let A ∈ L(V, V 0 ) be a self-adjoint, strictly positive operator. Then, A admits a sequence (λk )k∈N of real, strictly positive eigenvalues, each having finite multiplicity. The eigenvalues can be ordered into a nondecreasing sequence, such that lim λk = +∞ ,

k→+∞

(A.26)

and the corresponding eigenvectors form a complete orthonormal system in H. PROOF By theorem A.32, A−1 is, as an operator in H, linear, continuous, selfadjoint and compact. A−1 is also strictly positive, since so is A. Indeed, let u ∈ H, and set v := A−1 u ∈ H. Since in fact v ∈ V, and Av = u ∈ H, it follows that v ∈ dom(A). Then, hA−1 u, uiH = hv, AviH > 0 . By theorem A.36, A−1 admits a system of eigenvectors (w j ) j∈N corresponding to eigenvalues (µ j ) j∈N ∈ ]0, +∞[, with µ j → 0 as j → +∞. From the identities A−1 w j = µ j w j ,

j ∈ N,

340

Appendix: Selected Results from Analysis

setting λ j :=

1 µj

we deduce that Aw j = λ j w j ,

j ∈ N;

(A.27)

that is, each λ j is an eigenvalue of A, relative to the same eigenvector w j . Clearly, (A.26) follows from (A.25). The rest of the proof is immediate. We remark that if in addition A is coercive, we can endow V with the equivalent inner product defined by V × V 3 (u, v) 7→ ((u, v))V := hAu, viV 0 ×V . Indeed, (A.23) implies that for all u ∈ V, α kuk2V ≤ hAu, uiV 0 ×V ≤ kAkL kuk2V . With respect to this choice of inner product, the system of eigenvectors (w j ) j∈N is orthogonal also in V. Indeed, for each j, k ∈ N we have ((w j , wk ))V = hAw j , wk iV 0 ×V = λ j hw j , wk iV 0 ×V = λ j hw j , wk iH .

A.2.9 Fractional Powers of Positive Operators. In this section we define the fractional powers of strictly positive, self-adjoint, compact operators in the context of a Gelfand triple V ,→ H ,→ V 0 of Hilbert spaces. In this case, the existence of an orthonormal system of eigenvectors allows us to define the fractional powers of the operator by means of Fourier series expansions, generalizing (A.7). Let A ∈ L(V, V 0 ), and assume that all assumptions of theorem A.37 are satisfied. In particular, the injection V ,→ H is compact. Consider the orthonormal system of eigenvectors (w j ) j∈N of A, corresponding to the sequence of eigenvalues (λ j ) j∈N . For s ∈ R, we can define a linear, self-adjoint, unbounded operator As : H → H, in the following way. When s ≥ 0, we assign as the domain of As the set ∞

dom(As ) := u ∈ H :

∑ λ j2s |hu, w j iH |2 < +∞

;

(A.28)

j=0

this set is clearly a linear subspace of H. When s < 0, we define dom(As ) as the completion of H with respect to the norm !1/2 ∞

u 7→

∑ λ j2s |hu, w j iH |2

.

(A.29)

j=0

For u ∈ dom(As ), s ∈ R, we define ∞

As u :=

∑ λ js hu, w j iH w j ,

j=0

(A.30)

A.2

341

Linear Spaces and their Duals

the series converging in H. THEOREM A.38 In the above described assumptions, for each s ∈ R the space dom(As ) is a Hilbert space, with respect to the inner product defined by ∞

(u, v) 7→ hu, vis :=

∑ λ j2s hu, w j iH hv, w j iH .

j=0

For each s1 , s2 ∈ R, with s1 ≥ s2 , dom(As1 ) ,→ dom(As2 )

(A.31)

densely, and with compact injection. The operator As1 −s2 is an isomorphism from dom(As1 ) into dom(As2 ). In particular, if s ≥ 0, each space dom(As ) is dense in H; 0 dom(A−s ) := (dom(As )) (that is, the topological dual of the domain of As , defined in −s s ∗ (A.28)), and A = (A ) (that is, A−s is the adjoint of As , as introduced in definition A.29). PROOF Most of the claims are immediate; see also Zeidler, [Zei95, sct. 5.8]. To show the compactness of the injection (A.31), consider a sequence (um )m∈N , bounded in dom(As1 ). Then, there is a subsequence (umk )k∈N , converging weakly to some element u in dom(As1 ). By replacing umk with umk − u, we can assume that u = 0. We have to show that umk → 0 in dom(As2 ) strongly, that is (recalling the definition (A.29) of the norm in dom(As )), that for all η > 0 there is K > 0 such that for all k ≥ K, ∞

2

∑ λ j2s2 humk , w j iH

≤η.

(A.32)

j=0

For j0 ∈ N>0 to be determined, split ∞

2

∑ λ j2s2 humk , w j iH

j0 −1

=

j=0

2

∑ λ j2s2 humk , w j iH

∞

+

{z

} |

=:S0k

. (A.33)

j= j0

j=0

|

2

∑ λ j2s2 humk , w j iH {z

}

=:R0k

Since the sequence (umk )k∈N is bounded in dom(As1 ), we can estimate ∞

R0k =

−2(s1 −s2 )

∑ λj

j= j0

−2(s1 −s2 )

≤ λ j0

2 −2(s −s ) 2s λ j 1 humk , w j iH ≤ λ j0 1 2

kumk k2dom(As1 )

−2(s1 −s2 )

≤ Mλ j0

,

∞

2

∑ λ j2s1 humk , w j iH

j= j0

(A.34)

for suitable constant M independent of k and j. Since λ j → +∞ as j → +∞, it follows from (A.34) that we can fix j0 ∈ N>0 such that for all k ∈ N, R0k ≤ 21 η .

(A.35)

342

Appendix: Selected Results from Analysis

Since the subsequence (umk )k∈N converges weakly to 0 in dom(As1 ), we have that for each j = 0, . . . , j0 − 1, ∞

2s

2s

humk , w j idom(As1 ) = ∑ λi 1 humk , wi iH hwi , w j iH = λ j 1 humk , w j iH → 0 i=0

as k → +∞. Consequently, also hum , w j iH 2 → 0 , k so that we can determine K > 0 such that for all k ≥ K, hum , w j iH 2 ≤ k

η s . 2 j0 λ j02

Then, for k ≥ K we have that 2s

S0k ≤ λ j0 2

j0 −1

2

∑ humk , w j iH

j=0

η

2s

≤ λ j0 2

2s 2 j0 λ j0 2

j0 =

1 η. 2

From this and (A.35), (A.33), we deduce that (A.32) holds. This completes the proof of the compactness of the injection (A.31). In particular, for s = 0, s = dom(A0 ) = H ,

1 2

and s = 1, we have that

dom(A1/2 ) = V ,

dom(A1 ) = dom(A) ,

(A.36)

as sets and as Hilbert spaces. In particular, the first of (A.36) follows from (A.7). REMARK A.39 A definition of fractional powers of positive operators can also be given in the more general case of a self-adjoint linear operator A in a separable Hilbert space H, with dense domain dom(A), under the assumption that A is COERCIVE in H, i.e. such that (compare to definition A.33) there is α > 0 such that for all x ∈ dom(A), hAu, uiH ≥ αkuk2X . We refer e.g. to Dautray-Lions, [DL90, sct. VIII.6].

A.2.10 Interpolation Spaces The theory of interpolation describes the construction, starting from two normed spaces X and Y, of a family of “intermediate” normed spaces Zθ , parametrized by θ ∈ [0, 1], so that Z0 := X ∩ Y ,→ Zθ ,→ Y =: Z1 .

(A.37)

A.2

Linear Spaces and their Duals

343

These spaces have, in particular, the property that for all x ∈ X ∩ Y, kxkZθ ≤ Ckxk1X−θ kxkθY ,

(A.38)

with C independent of x. The spaces Zθ are called INTERPOLATION SPACES between X ∩ Y and Y, and are denoted by Zθ = [X ∩ Y, Y]θ ; inequality (A.38) is called the corresponding INTERPOLATION INEQUALITY. For an overview of the general theory of the interpolation spaces, we refer e.g. to Bergh-Löfström, [BL76]; here, we limit ourselves to follow Lions-Magenes, [LM72, sct. 1.2], and recall one possible construction of interpolation spaces, starting from two separable Hilbert spaces X and Y. Since this construction rests heavily on the notion of the fractional powers of a positive operator, we assume that X ,→ Y densely, with continuous and compact injection; note, however, that, by remark A.39, the assumption of compactness is not required. Given X and Y as above, we consider X ,→ Y ,→ X 0 as a Gelfand triple, and define a linear, unbounded operator A : X → X 0 as follows. First, we assign as its domain the set of all u ∈ X such that the linear map X 3 v 7→ hu, viX ∈ R

(A.39)

is continuous on X with respect to the (weaker) topology induced by Y. Thus, if u ∈ dom(A), (A.39) defines a linear operator A : X → X 0 , by hAu, viX 0 ×X = hu, viX ,

v∈X.

(A.40)

As in section A.2.7, we can then consider A as an unbounded linear operator in Y, with domain dom(A) dense in Y. Moreover, (A.40) implies that A is self-adjoint and strictly positive; in fact, for all u ∈ dom(A), hAu, uiY = hAu, uiX 0 ×X = hu, uiX = kuk2X . Consequently, for s ≥ 0 we can define, as in section A.2.9, the fractional powers As of A, as linear, unbounded operators in Y. In particular, the operator Λ := A1/2 is also positive and self-adjoint; moreover, recalling (A.36), dom(Λ ) = X and, by (A.40), for all u, v ∈ X , hu, viX = hΛ u,Λ viY .

(A.41)

We define then, for θ ∈ [0, 1], the interpolation spaces [X , Y]θ := dom(Λ 1−θ ) .

(A.42)

These spaces are Banach spaces, with respect to the graph norm defined by kuk2θ := kuk2Y + kΛ 1−θ uk2Y . In particular, note that [X , Y]0 = dom(Λ ) = X , in accord with (A.37).

[X , Y]1 = dom(IY ) = Y ,

(A.43)

344

Appendix: Selected Results from Analysis

A.2.11 Differential Calculus in Banach Spaces In this section we recall some basic results concerning homeomorphisms and diffeomorphisms between Banach spaces. DEFINITION A.40 Let X and Y be Banach spaces, and U ⊆ X be open. Consider a map F : U → Y, and let x ∈ U. F is said to be (Fréchet) DIFFERENTIABLE at x, if there exists an operator A ∈ L(X , Y) such that lim

h∈X h→0

kF(x + h) − F(x) − AhkY = 0. khkX

It is easily seen that A is uniquely determined by x (see e.g. Ambrosetti-Prodi, [AP93, sct. 1.1]); we call A the F RÉCHET derivative of F at x, and write A =: F 0 (x). In particular, note that if F is linear and continuous (i.e., F ∈ L(X , Y)), then F is differentiable at each x ∈ X , with F 0 (x) = F (this is because F(x+h)−F(x) = F(h)). DEFINITION A.41 In the same conditions of definition A.40, assume that F is differentiable at each x ∈ U. The map U 3 x 7→ F 0 (x) ∈ L(X , Y) is called the DERIVATIVE of F. If F 0 is continuous, we say that F is CONTINUOUSLY DIFFERENTIABLE in U, and write F ∈ C1 (U, Y). DEFINITION A.42 Let X and Y be normed linear spaces, and U ⊆ X , V ⊆ Y. A map F : U → V is called a HOMEOMORPHISM, if F is a continuous bijection, and its inverse F −1 : V → U is also continuous. If in addition both X and Y are Banach spaces and both F and F −1 are continuously differentiable, F is called a DIFFEOMORPHISM . DEFINITION A.43 A mapping F : D ⊆ V → V in a normed, linear space V is called NORM - COERCIVE if for any γ ≥ 0 there exists a closed, bounded set Dγ ⊆ D such that kFyk > γ for all y ∈ D \ Dγ . (Compare e.g. with Rheinboldt, [Rhe69, def. 3.6].) THEOREM A.44 Let V be a normed linear space and G : D ⊆ V → V a sequential compact mapping on the P(V)-path-connected1 set D ⊆ V. Suppose further that F = I − G is a normcoercive local homeomorphism and that F(D) is open. Then G is a bijection, and hence a homeomorphism, from D onto V. 1 In

[Rhe69], P(V) denotes the set of all continuous paths in V.

A.3

Semigroups of Linear Operators

345

PROOF See e.g. [Rhe69, thm. 3.7] The following theorem describes a sufficient condition for a differentiable map to be a local diffeomorphism. THEOREM A.45 Let X and Y be Banach spaces, F ∈ C1 (X , Y), and x ∈ X be such that F 0 (x) is invertible (as a linear map from X into Y). Then F is locally invertible at x, with a C1 inverse. More precisely, there are neighborhoods U of x in X , and V of F(x) in Y, such that F is a diffeomorphism from U into V. Moreover, for all y ∈ V, the inverse differentiation formula −1 0 F −1 (y) = F 0 (F −1 (y)) holds. PROOF See e.g. Ambrosetti-Prodi, [AP93, sct. 2.1, thm. 1.2].

A.3 Semigroups of Linear Operators In this section we report the most fundamental results on semigroup theory, as relevant to their applications to semilinear evolution equations of the form (4.4). Most of the material we present is taken from Pazy, [Paz83]. As before, we assume that the spaces we consider are at least separable Banach spaces.

A.3.1 General Results 1. We start by introducing the definition of semigroups depending on a real parameter. DEFINITION A.46 Let S = (S(t))t ≥0 be a family of linear, continuous operators in X (i.e., S(t) ∈ L(X , X ) for all t ≥ 0). 1. S is said to be a SEMIGROUP if S(0) = IX

(A.44)

(the identity in X ), and for all t, θ ≥ 0, S(t + θ ) = S(t)S(θ ) = S(θ )S(t) .

(A.45)

2. A semigroup S is said to be CONTINUOUS if for each x ∈ X , the map t 7→ S(t)x is continuous from [0, +∞[ to X . In this case, S is also called a C0 -semigroup.

346

Appendix: Selected Results from Analysis

Given a semigroup S on X , we can define a linear operator A on X , in general unbounded, in the following way. First, we assign as its domain the subspace dom(A) := {x ∈ X : lim

t →0+

S(t)x − x 0 (0)x =: S+ t

exists in X } ;

then, we define A : dom(A) → X by 0 (0) . A := S+

It is clear that A is a linear operator; A is called the INFINITESIMAL GENERATOR (or, more simply, the GENERATOR) of the semigroup S. The following fundamental result, known as the H ILLE -YOSIDA THEOREM, relates the properties of a C0 -semigroup and its infinitesimal generator. THEOREM A.47 Let A : X → X be a linear operator, not necessarily bounded. A is the infinitesimal generator of a C0 -semigroup S, if and only if its domain dom(A) is dense in X , A is closed, the resolvent set ρ(A) (recall definition A.21) contains the interval [0, +∞[, and for all λ > 0, kR(A, λ )kL ≤

1 . λ

PROOF See e.g. Pazy, [Paz83, sct. 1.3]. 2. The definition of semigroup can be extended to families of linear continuous operators on X , depending on a complex parameter, which necessarily varies in an additive semigroup of the complex plane C. In particular, we will consider parameters varying in an open sector Σ := {z ∈ C : ϕ1 < arg z < ϕ2 , ϕ1 < 0 < ϕ2 } ,

(A.46)

containing the nonnegative real axis [0, +∞[. DEFINITION A.48 Let Σ be as in (A.46), and S = (S(z))z∈Σ be a family of linear, continuous operators in X . S is said to be an ANALYTIC SEMIGROUP if the following conditions are satisfied: 1. The map z 7→ S(z) is analytic from Σ to L(X , X ); 2. The semigroup properties (A.44) and (A.45) hold, i.e. if S(0) = IX and for all z, ζ ∈ Σ , S(z + ζ ) = S(z)S(ζ ) = S(ζ )S(z) ;

A.3

Semigroups of Linear Operators

347

3. For each x ∈ X , the map z 7→ S(z)x is continuous from Σ to X . The following result extends to analytic semigroups the characterization of C0 semigroups given by the Hille-Yosida theorem A.47. THEOREM A.49 Let A : X → X be the infinitesimal generator of a semigroup S. Then, S is analytic (that is, S can be extended to an analytic semigroup defined on a sector Σ as in (A.46)) if and only if there are positive constants C and Λ such that for all n ∈ N>0 and all λ > nΛ , kA R(A, λ )n+1 kL ≤

C . nλ n

PROOF See e.g. Pazy, [Paz83, sct. 2.5]. 3. Definition A.46 can be naturally extended to that of a GROUP S = (S(t))t ∈R of linear, continuous operators on X , by requiring that all statements in the definition be valid for all t, θ ∈ R. In particular, we have the following characterization of unitary groups (see definition A.20) in Hilbert spaces, known as S TONE ’ S THEOREM. THEOREM A.50 Let X be a Hilbert space, and A : X → X be a linear operator, with domain dom(A) dense in X . Then, A is the generator of a unitary group S on X if and only if A∗ = −A. PROOF See e.g. Engel-Nagel, [EN00, sct. 2.3]. We recall that if X is a linear space on C, with imaginary unit i, the condition A∗ = −A is equivalent to the requirement that iA be self-adjoint (because ¯i = −i).

A.3.2 Applications to PDEs The theory of semigroups allows us to solve the initial value problem for evolution equations of the form (3.14), i.e. Ut + AU = F(U) ,

(A.47)

by interpreting them as abstract ODEs in a Banach space X . DEFINITION A.51 Let A : X → X be a linear operator, not necessarily bounded. Let U0 ∈ X , and T > 0.

348

Appendix: Selected Results from Analysis

1. A function U ∈ C([0, T ]; X ) is a MILD SOLUTION of the initial value problem for (A.47), with initial value U(0) = U0 , if U satisfies the integral equation U(t) = e−tAU0 +

Z t

e−(t −θ )A F(U(θ )) dθ ,

0≤t ≤T .

0

2. A function U ∈ AC([0, T ]; X ) is a STRONG SOLUTION of the same IVP, if equation (A.47) is satisfied for almost all t ∈ [0, T ]. (Recall that, by theorem A.4, U is differentiable almost everywhere in [0, T ].) THEOREM A.52 Assume that A is the generator of a C0 -semigroup on X , and that F : X → X is globally Lipschitz continuous. Then for all U0 ∈ X , the initial value problem for (A.47), with initial value U(0) = U0 , has a unique mild solution U. Moreover, for all T > 0 the map X 3 U0 7→ U ∈ C([0, T ]; X ) is Lipschitz continuous. If in addition U0 ∈ dom(A), then U is a strong solution. PROOF See e.g. Pazy, [Paz83, sct. 6.1]. Theorem A.52 can be applied directly to evolution equations of “parabolic” type, such as the heat equation (3.1), where A = −∆, which we know to generate an analytic semigroup on L2 (Ω ) (see theorem A.79 below). This is not the case for the wave equation (3.4), which has first to be converted into an equivalent first order system of the form (3.14). Recalling (5.113) and (5.114), the operator A has then the matrix form A=

0 −∆

− ε1 1 ε

! ,

(A.48)

and X is the product space H10 (Ω ) × L2 (Ω ). In this case, we resort instead to the following result: THEOREM A.53 Let ε > 0, and A be as in (A.48). Then, A generates a C0 -semigroup on X = H10 (Ω ) × L2 (Ω ). PROOF See e.g. Pazy, [Paz83, sct. 7.2].

A.4

Lebesgue Spaces

349

A.4 Lebesgue Spaces In this section we review the main properties of the Lebesgue spaces L p (Ω ), 1 ≤ p ≤ +∞. Here, Ω denotes an arbitrary domain of RN ; we allow Ω = RN . We consider in Ω the standard Lebesgue measure, and integration in Ω is meant in the Lebesgue sense. When we do not provide a reference for the proof of a result, such proof can be found e.g. in Adams, [Ada78, ch. 2].

A.4.1 The Spaces L p (Ω ) DEFINITION A.54 Let p ∈ [1, +∞]. 1. If p 6= +∞, L p (Ω ) is the space of all equivalence classes, with respect to the equivalence relation f ∼ g ⇐⇒ meas{x ∈ Ω : f (x) 6= g(x)} = 0 , of the measurable functions f : Ω → R, such that Z

| f (x)| p dx < +∞.

Ω

2. If p = +∞, L∞ (Ω ) is the space of all equivalence classes of the measurable functions f : Ω → R which are ESSENTIALLY BOUNDED; that is, def.

f ∈ L∞ (Ω ) ⇐⇒ ∃ M > 0 : | f (x)| ≤ M a.e. in Ω .

(A.49)

As usual, with abuse of notation we identify an equivalence class with any one of its representatives, which we still call a “function”. p (Ω ) of locally p-integrable (or, if p = +∞, 3. Finally, we define the space Lloc locally bounded) functions by def. p (Ω ) ⇐⇒ (∀ K ⊂ Ω , K compact : f K ∈ L p (K)) . f ∈ Lloc

Each L p (Ω ) is a linear space, which can be endowed with the norms 1/p Z p |u(x)| dx |u| p := if 1 ≤ p < +∞ ,

(A.50)

Ω

|u|∞ := supessx∈Ω |u(x)| := inf{M > 0 : (A.49) holds } .

(A.51)

We list the major properties of these spaces in the following theorem. THEOREM A.55 Let 1 ≤ p ≤ +∞, and consider in L p (Ω ) the norms defined in (A.50) and (A.51). Then:

350

Appendix: Selected Results from Analysis

1. L p (Ω ) is a Banach space; 2. L2 (Ω ) is a Hilbert space, with respect to the scalar product h f , gi :=

Z

f (x)g(x) dx

or

h f , gi :=

Ω

Z

f (x)g(x) dx ,

(A.52)

Ω

depending respectively on whether the underlying scalar field is R, or C; 3. If 1 ≤ p < +∞, the space C∞ 0 (Ω ) of the infinitely differentiable functions with support compact in Ω is dense in L p (Ω ); 4. If 1 ≤ p < +∞, L p (Ω ) is separable, while L∞ (Ω ) is not; 5. If 1 < p < +∞, L p (Ω ) is reflexive; neither L1 (Ω ) nor L∞ (Ω ) are reflexive. We can also give a definition of the spaces L p (Γ ), where Γ is a sufficiently smooth (N − 1)-dimensional submanifold of RN (see e.g. definition 5.1 of chapter 5). This is done in a natural way, by means of charts of local coordinates, which are required to have an image in L p (RN ).

A.4.2 Inequalities 1 We denote by p, q etc. generic numbers in [1, +∞], set ∞ := 0, and call p and q if 1 1 + = 1. p q In particular, the pairs (p, q) = (2, 2), (p, q) = (1, +∞), and (p, q) = (+∞, 1) are conjugate indices.

CONJUGATE INDICES

PROPOSITION A.56 Let p, q ∈ ]1, +∞[ be conjugate indices. Then, YOUNG ’ S INEQUALITY ab ≤

1 p 1 q a + b p q

(A.53)

holds, for all a, b ≥ 0. More generally, for all η > 0 there is Cη > 0 such that ab ≤ ηa p +Cη bq .

(A.54)

In particular, for p = q = 2, (A.53) and (A.54) read 1 1 1 2 ab ≤ a2 + b2 , ab ≤ ηa2 + b . 2 2 4η THEOREM A.57 Let p, q ∈ [1, +∞] be conjugate indices. Then for all u ∈ L p (Ω ) and v ∈ Lq (Ω ), the product uv ∈ L1 (Ω ), and H ÖLDERS ’ S INEQUALITY |uv|1 ≤ |u| p |v|q

(A.55)

A.4

Lebesgue Spaces

351

holds. More generally, for all η > 0 there is Cη > 0 such that |uv|1 ≤ η|u| pp +Cη |v|qq . As a consequence, M INKOWSKI ’ S INEQUALITY holds, for u, v ∈ L p (Ω ): |u + v| p ≤ |u| p + |v| p . For functions in the Hilbert space L2 (Ω ), Hölder’s inequality (A.55) reads h f , gi ≤ | f |2 |g|2 , and is usually known as S CHWARZ ’ INEQUALITY. Hölder’s inequality (A.55) can be generalized to the product of any finite number of functions. THEOREM A.58 Let r and p1 , . . . , pk ∈ [1, +∞] be such that k

∑

j=1

1 1 = . pj r

(A.56)

Assume f j ∈ L p j (Ω ), for 1 ≤ j ≤ k. Then the product f1 · · · fk is in Lr (Ω ), and satisfies the estimate | f1 · · · fk |r ≤ | f1 | p1 · · · | fk | pk .

(A.57)

PROOF The result can be easily proven by induction on k.

A.4.3 Other Properties of the Spaces L p (Ω ) 1. The Dual of L p (Ω ). THEOREM A.59 Let p ∈ [1, +∞[, and q be its conjugate index. Then, the topological dual of L p (Ω ) is isometrically isomorphic to the space Lq (Ω ). Consequently, if p ∈ ]1, +∞[ and ( fm )m∈N ⊂ L p (Ω ), then fm → f weakly in L p (Ω ) if and only if for all g ∈ Lq (Ω ), Z

fm (x)g(x) dx →

Ω

Z

f (x)g(x) dx . Ω

Analogously, fm → f weakly∗ in L∞ (Ω ) if and only if for all g ∈ L1 (Ω ), Z Ω

fm (x)g(x) dx →

Z

f (x)g(x) dx . Ω

352

Appendix: Selected Results from Analysis

2. (L p (Ω )) p≥1 as a Family of Interpolation Spaces. If p < q < r, it is natural to expect the space Lq (Ω ) to be an “intermediate” space between Lr (Ω ) and L p (Ω ). This is indeed the case if Ω has finite measure, in which case actually Lr (Ω ) ,→ Lq (Ω ) ,→ L p (Ω ). PROPOSITION A.60 Let Ω have finite measure |Ω |. If 1 ≤ p ≤ q ≤ +∞, then Lq (Ω ) ,→ L p (Ω ), with continuous imbedding. More precisely, for all u ∈ Lq (Ω ), 1

1

|u| p ≤ |Ω | p − q |u|q . More generally, we can characterize the spaces L p (Ω ) as a family of interpolation spaces (see section A.2.10). THEOREM A.61 Let p, r ∈ [1, +∞], with p ≤ r. For 0 ≤ λ ≤ 1, define q by 1 λ 1−λ = + . q p r

(A.58)

Then q ∈ [p, r], with q = p if λ = 1 and q = r if λ = 0, and Lq (Ω ) = [Lr (Ω ) ∩ L p (Ω ), L p (Ω )]λ . That is, if u ∈ L p (Ω ) ∩ Lr (Ω ), then u ∈ Lq (Ω ), and satisfies the interpolation inequality |u|q ≤ |u|λp |u|1r −λ .

(A.59)

PROOF See Bergh-Löfström, [BL76, ch. 5]. Note that (A.59) is a consequence of Hölder’s inequality: indeed, letting α = λ q and β = (1 − λ )q, we easily verify that (A.58) implies that a = αp and b = βr are conjugate indices.

A.5 Sobolev Spaces of Scalar Valued Functions In this section we review the main properties of the Sobolev spaces Hs (Ω ), s ∈ R. Here, we assume that either Ω = RN , or that Ω is bounded. In the latter case, we assume that its boundary ∂ Ω is sufficiently smooth; more precisely, that ∂ Ω is an (N − 1)-dimensional submanifold of RN , of class Cm , for suitable m ≥ 0. As in section A.4 we consider in Ω the Lebesgue measure, and integration is meant in the Lebesgue sense. Proof of the results for which we do not provide a reference can be found e.g. in Adams, [Ada78, chs. 3-6], or in Lions-Magenes, [LM72, ch. 1].

A.5

Sobolev Spaces of Scalar Valued Functions

353

A.5.1 Distributions in Ω The linear space C∞ 0 (Ω ) of the infinitely differentiable functions with compact support in Ω can be endowed with a locally convex (but not metrizable) topology (see e.g. Rudin, [Rud73, ch. 6]). The corresponding topological space is denoted by D(Ω ), and called the space of TEST FUNCTIONS. The space of linear continuous functionals on D(Ω ), called DISTRIBUTIONS, is denoted by D0 (Ω ). We also denote by h · , · iD(Ω ) the duality product between D0 (Ω ) and D(Ω ). Given α := (α1 , . . . , αN ) ∈ NN , we call α a MULTIINDEX, define its LENGTH as |α| := α1 + · · · + αN , and set Dα :=

∂ |α | α α ∂ x1 1 · · · ∂ xNN

.

Any function f ∈ L2 (Ω ) defines a distribution F, by hF, ϕiD(Ω ) :=

Z

f (x)ϕ(x) dx ;

(A.60)

Ω

to denote the dependence of F on f , we write F := T f . The converse, however, is not true in general; for example, for the so-called D IRAC δ distribution, defined by hδ , ϕiD(Ω ) := ϕ(0) , there is no f ∈ L2 (Ω ) such that δ = T f . This fact motivates the following definition. DEFINITION A.62 A distribution F is REGULAR, if there is f ∈ L2 (Ω ) such that F = Tf . The importance of distributions in the theory of PDEs resides in the fact that any distribution F ∈ D0 (Ω ) can be differentiated in distributional sense. That is, given any multiindex α, we can define a new distribution Dα F, called the DISTRI BUTIONAL DERIVATIVE of F of order α, by the identity hDα F, ϕiD(Ω ) := (−1)|α | hF, Dα ϕiD(Ω ) ,

ϕ ∈ D(Ω ) .

(A.61)

This definition is natural, in the sense that if a distribution F is regular, i.e. F = T f , and f ∈ Cm (Ω ), then we can integrate the right side of (A.61) by parts, and deduce that Dα F = TDα f for all α with |α| ≤ m.

A.5.2 The Spaces Hm (Ω ), m ∈ N The spaces Hm (Ω ), m ∈ N, consist of subspaces of L2 (Ω ), whose functions possess weak derivatives of all orders up to m. More precisely, note that, given any function f ∈ L2 (Ω ), the distribution T f can be differentiated in distributional sense, as in (A.61). Consequently, we can give

354

Appendix: Selected Results from Analysis

DEFINITION A.63 Let f ∈ L2 (Ω ), and α ∈ NN . We say that f has a WEAK DERIVATIVE f α ∈ L2 (Ω ) if the distribution Dα T f is regular; that is, if there is f α ∈ L2 (Ω ) such that Dα T f = T fα . In this case, with abuse of notation we write that Dα T f ∈ L2 (Ω ). This definition A.63 makes sense, since it is easily seen that fα is uniquely determined (up to a set of measure zero in Ω ) by the distribution Dα T f . Recalling (A.61) and (A.60), definition A.63 means that fα ∈ L2 (Ω ) is the weak derivative of order α of a function f ∈ L2 (Ω ) if for all ϕ ∈ dom(Ω ), Z

fα (x)ϕ(x) dx = (−1)|α |

Z

f (x)(Dα ϕ)(x) dx .

(A.62)

Ω

Ω

Note that (A.62) generalizes the classical integration by parts formula, which would hold if f ∈ C|α | (Ω ); indeed, in this case fα = Dα f , and the vanishing of ϕ in a neighborhood of ∂ Ω causes the absence of the boundary terms in (A.62). DEFINITION A.64 Let m ∈ N. We set Hm (Ω ) := {u ∈ L2 (Ω ) : (∀ α ∈ NN , |α| ≤ m : Dα Tu ∈ L2 (Ω ))} . With abuse of notation, we set Dα Tu =: Dα u. Thus, if m ≥ 1, functions in Hm (Ω ) have weak derivatives of order up to m in This notion of differentiability, however, obviously does not coincide with the classical one. In particular, neither of the spaces Hm (Ω ) and Cm (Ω ) is in general contained in the other. Each Hm (Ω ) is a linear space, which can be endowed with the norm !1/2 Z

L2 (Ω ).

kukm :=

|Dα u(x)|2 dx

∑

.

|α |≤m Ω

In particular, H0 (Ω ) = L2 (Ω ) (as a set and as a normed space). We list the major properties of these spaces in the following theorem. THEOREM A.65 For each m ∈ N, Hm (Ω ) is a separable Banach space, in which the subspace C∞ (Ω )∩ Hm (Ω ) is dense. In fact, Hm (Ω ) is a Hilbert space, with respect to the scalar product h f , gim :=

Z

∑

(Dα f )(x)(Dα g)(x) dx

|α |≤m Ω

(with modification analogous to the second of (A.52) if the underlying scalar field is C). s REMARK A.66 In general, it is not true that C∞ 0 (Ω ) is dense in H (Ω ); this does N hold, however, if m = 0, or if Ω = R .

A.5

Sobolev Spaces of Scalar Valued Functions

355

A.5.3 The Spaces Hs (Ω ), s ∈ R≥0 The spaces Hs (Ω ), for general s ∈ [0, +∞[, can be defined by interpolation. More precisely, given s ≥ 0, we choose m ∈ N such that 0 ≤ s ≤ m, and set Hs (Ω ) := [Hm (Ω ), L2 (Ω )]1−s/m .

(A.63)

The following result shows that this definition is consistent. THEOREM A.67 The space Hs (Ω ) defined in (A.63) is independent, up to norm equivalence, of the choice of m ≥ s. Moreover, Hs (Ω ) can be defined by interpolation between consecutive integers; that is, setting m := bsc (the integer part of s), again up to norm equivalence we have that Hs (Ω ) := [Hm+1 (Ω ), Hm (Ω )]m+1−s .

(A.64)

PROOF See e.g. Lions-Magenes, [LM72, sct. 1.9]. REMARK A.68 When s ∈ N, definition (A.64) is in accord with the definition given in A.64. Indeed, if s ∈ N, then m = s, so that, by (A.43), Hs (Ω ) = [Hm+1 (Ω ), Hm (Ω )]1 = Hm (Ω ) .

As a consequence of general results of interpolation theory, the spaces Hs (Ω ) inherit the same properties listed in theorem A.65. The following theorem gives some additional properties of the spaces Hs (Ω ). THEOREM A.69 For s ≥ 0, let Hs (Ω ) be defined as in (A.63). Then: 1. Hs (Ω ) as interpolation spaces. For s1 and s2 such that 0 ≤ s2 ≤ s1 , and θ ∈ [0, 1], we have [Hs1 (Ω ), Hs2 (Ω )]θ = H(1−θ )s1 +θ s2 . 2. Imbeddings of Hs (Ω ). The following imbeddings are continuous: 2.1. If 2s < N, Hs (Ω ) ,→ Lq (Ω ), for 2 ≤ q ≤

2N N −2s .

2.2. If 2s = N, Hs (Ω ) ,→ Lq (Ω ), for 2 ≤ q < +∞. 2.3. If 2s > N, Hs (Ω ) ,→ Cb (Ω ), where Cb (Ω ) := C(Ω ) if Ω is bounded, while Cb (RN ) denotes the subspace of the functions in C(RN ) which are bounded.

356

Appendix: Selected Results from Analysis If in addition Ω is bounded and s > 0, the imbedding Hs (Ω ) ,→ Hs−ε (Ω ) is compact for all ε ∈ ]0, s].

3. Hs (Ω ) as an algebra. If 2s > N, Hs (Ω ) is an algebra; that is, if f and g ∈ Hs (Ω ), their pointwise product f · g (which makes sense, because f and g are continuous, by part 2.2 above) is also in Hs (Ω ). PROOF In addition to Adams, [Ada78, chs. 3-6], or in Lions-Magenes, [LM72, ch. 1]. For the imbeddings when s is not an integer, see also Triebel, [Tri95, scts. 2.3, 2.6], or Peetre, [Pee66]. The imbeddings described in theorem A.69 are a consequence of the following more general result, due to Gagliardo and Nirenberg. THEOREM A.70 Let m ∈ N, p, r ∈ [1, +∞], and u ∈ Hm (Ω ) ∩ Lr (Ω ). For integer j ≤ m, and θ ∈ [ mj , 1] (with the exception θ 6= 1 if m − j − N2 ∈ N), define q by j 1 = +θ q N

1 1 m − + (1 − θ ) . 2 N r

(A.65)

Then for any γ ∈ NN , with |γ| = j, Dγ u ∈ Lq (Ω ), and satisfies the G AGLIARDO N IRENBERG inequality kDγ ukLq (Ω ) ≤ C

∑

θ kDα ukθL2 (Ω ) kuk1L− r (Ω ) +C1 kukLs (Ω ) ,

(A.66)

|α |=m

where s := max(2, r), and C > 0, C1 ≥ 0 are independent of u. The choice C1 = 0 is admissible if Ω = RN . PROOF See e.g. Racke, [Rac92, ch. 4]. REMARK A.71 1. The choice j = 0, θ = 1 in (A.65) yields the original Sobolev 2N imbedding Hm (Ω ) ,→ Lq (Ω ), with q = N − 2m , described in part (2.1) of theorem A.69 for integer s = m. The choice of the lowest possible value θ = mj in (A.65) yields the dimensionless version of estimate (A.66) kDγ ukLq (Ω ) ≤ C

∑

|α |=m

j/m

1− j/m

kDα ukL2 (Ω ) kukLr (Ω ) +C1 kukLs (Ω ) ,

with |γ| = j and q ≥ 1 defined by j 1 j 1 = + 1− . q 2m r m

A.5

357

Sobolev Spaces of Scalar Valued Functions

2. Unless u satisfies some extra conditions, such as having vanishing “trace” at ∂ Ω (see section A.5.4 below), it is not in general possible to choose C1 = 0 in (A.66) if Ω 6= RN . To see this, it is sufficient to consider the example u(t) = t in Ω = ]0, 1[⊂ R, with m = 2, j = 1, θ = 21 and r = q = 2: if (A.66) held with C1 = 0, we would deduce the contradiction u0 (t) ≡ 0. We recall that when Ω = RN , the spaces Hs (RN ) can be defined in an alternative way, by means of the Fourier transform. The two definitions coincide, up to isomorphisms. In particular, we can then define, by means of charts of local coordinates, the Sobolev spaces Hs (Γ ), where Γ is a (N − 1)-dimensional submanifold of RN of class Cm , m ≥ s.

A.5.4 The Spaces Hs0 (Ω ), s ∈ R≥0 , and Hs (Ω ), s ∈ R<0 We now assume that Ω is bounded, and that its boundary is, as already stated, a (N − 1)-dimensional submanifold of RN of class Cm , m ≥ 0. Given a function f ∈ Cm (Ω ), we can then define its TRACES at ∂ Ω γ j f :=

∂jf , ∂ν j

0 ≤ j ≤ m,

(A.67)

where ν denotes the unit outward normal vector to ∂ Ω . If m ≥ 1 and s ∈ ] 21 , m], we can extend the definition of these traces to functions in Hs (Ω ), by means of a density argument. THEOREM A.72 Let Ω be as stated above, and 0 ≤ s ≤ m. The space Cm (Ω ) is dense in Hs (Ω ). THEOREM A.73 Let Ω be as stated above, and assume that m ≥ s > 21 . Let m0 be the largest integer such that 0 ≤ m0 < s − 12 . Then for 0 ≤ j ≤ m0 , the TRACE OPERATOR γ j defined in (A.67) extends by continuity to a linear, continuous map from Hs (Ω ) to Hs− j−1/2 (∂ Ω ). The map m0

γ˜ : Hs (Ω ) 7→ ∏ Hs− j−1/2 (∂ Ω )

(A.68)

j=0

is surjective, and has a continuous inverse. PROOF See e.g. Lions - Magenes, [LM72, sct. 1.9]. In particular, functions in H1 (Ω ) have a trace in H1/2 (∂ Ω ). For s ≥ 0 we define the subspace Hs0 (Ω ) of Hs (Ω ) as the closure of C∞ 0 (Ω ) in s H (Ω ).

358

Appendix: Selected Results from Analysis

Since the map γ˜ defined in (A.68) is surjective, and vanishes on C∞ 0 (Ω ), it follows that this space cannot be dense in Hs (Ω ) if s > 21 . Hence, in this case, Hs0 (Ω ) is a proper subspace of Hs (Ω ) (unless Ω = RN ). THEOREM A.74 1 s s The space C∞ 0 (Ω ) is dense in H (Ω ) if and only if s ≤ 2 , in which case, H0 (Ω ) = Hs (Ω ). If s > 21 , Hs0 (Ω ) = {u ∈ Hs (Ω ) : γ j (u) = 0 , 1 ≤ j < s − 12 } . Furthermore, if s − 12 ∈ / N, and m ∈ N is such that m ≥ s, 2 Hs0 (Ω ) = [Hm 0 (Ω ), L (Ω )]1−s/m .

(A.69)

PROOF See e.g. Lions-Magenes, [LM72, ch. 11]; in particular, (A.69) is contained in their theorem 11.6, since H00 (Ω ) = L2 (Ω ) by the first claim of theorem A.74. Finally, for s < 0 we set 0 s Hs (Ω ) := H− 0 (Ω ) (topological dual). Thus, in particular, H−1 (Ω ) is defined to be the dual of H10 (Ω ). We conclude with the following interpolation theorem. THEOREM A.75 Let s1 , s2 ≥ 0, with s2 6= m + 21 , m ∈ N. Then, for θ ∈ [0, 1], [Hs1 (Ω ), H−s2 (Ω )]θ = H(1−θ )s1 −θ s2 (Ω ) , provided that (1 − θ )s1 − θ s2 6= n + 21 , n ∈ N. PROOF See e.g. Lions-Magenes, [LM72, sct. 1.12.4].

A.5.5 The Laplace Operator In chapter 3, we have considered the standard example of the Gelfand triple consisting of the Sobolev spaces V := H10 (Ω ) ,→ H := L2 (Ω ) ,→ V 0 = H−1 (Ω ) , with Ω ⊂ Rn a bounded domain. In particular, the boundedness of Ω implies that the injection H10 ,→ L2 (Ω ) is compact (by part 2 of theorem A.69). In the sequel, we denote as usual by k · km the norm in Hm (Ω ), and abbreviate k · k0 = k · k.

A.5

359

Sobolev Spaces of Scalar Valued Functions

1. The operator A := −∆ is clearly in L(V, V 0 ); as such, A is self-adjoint and strictly positive, as we see from the identities h−∆u, viV 0 ×V = h∇u, ∇viH = h−∆v, uiV 0 ×V , h−∆u, uiV 0 ×V = h∇u, ∇uiH = k∇uk2 .

(A.70)

Moreover, by theorem A.32, A has compact inverse. Consequently, from theorem A.37 we deduce THEOREM A.76 The operator A = −∆ admits an unbounded sequence of positive eigenvalues (λ j ) j∈N . This sequence can be ordered so that 0 < λ1 < λ2 ≤ · · · ≤ λ j ≤ · · · ,

λ j → +∞ .

(A.71)

For each j ∈ N, the corresponding eigenvector w j is in C∞ (Ω ) ∩ C(Ω ), and the sequence (w j ) j∈N is a complete orthogonal system in L2 (Ω ). 2.

We now show that and that we can choose in V = H10 (Ω ) the norm u ∈ H10 (Ω ) .

kukV := k∇uk ,

(A.72)

Formally, this is a consequence of (A.27). Indeed, given u ∈ V, if we consider its Fourier series expansion (A.7), in terms of the total basis of the eigenvectors of −∆, recalling (A.71) and that these eigenvectors are orthonormal in L2 (Ω ), we have that * + * + ∞

k∇uk2 = h−∆u, ui =

j=0

∞

=

∞

−∆ ∑ α j w j , ∑ αk wk

∞

=

k=0

∞

∑ λ j α j w j , ∑ αk wk

j=0

k=0

∞

∑ λ j α 2j ≥ λ0 ∑ α 2j = λ0 kuk2 .

j=0

j=0

That is, for all u ∈ H10 (Ω ), k∇uk2 ≥ λ0 kuk2 ,

(A.73)

which is known as P OINCARÉ ’ S INEQUALITY. It follows that −∆ is well defined and bijective from V into V 0 . This is a consequence of Lax-Milgram’s theorem A.34, since A is self-adjoint, positive and coercive, as we see from (A.70). 3. As an operator in L2 (Ω ), −∆ has domain dom(A) = H2 (Ω ) ∩ H10 (Ω ) ; to see this, we need to recall the following elliptic regularity result.

360

Appendix: Selected Results from Analysis

THEOREM A.77 Consider the elliptic boundary value problem ( −∆u = f in Ω , u=0 on ∂ Ω .

(A.74)

Let f ∈ H−1 (Ω ), and u ∈ H10 (Ω ) be the corresponding unique solution of (A.74), whose existence is assured by the Lax-Milgram theorem. If in addition f ∈ L2 (Ω ), then u ∈ H2 (Ω ), and there is C > 0, independent of u, such that kuk2 ≤ C(k∆uk + kuk) . PROOF See e.g. Gilbarg-Trudinger, [GT83, sct. 8.4]. In conclusion, theorems A.37 and A.38 are applicable, and we can thus define the fractional powers (−∆)s , for all s ∈ R. 4. We now show that, if for s ≥ 0 we set ˜ s (Ω ) := dom((−∆)s/2 ) , H

(A.75)

then THEOREM A.78 ˜ s (Ω ) = Hs (Ω ). For all s ≥ 0, such that s − 21 ∈ / N, H 0 PROOF Given s ≥ 0, fix m ∈ N, with m ≥ s, and consider the operator 2 Am := (−∆)m : Hm 0 (Ω ) → L (Ω ) ,

as an unbounded operator with domain H2m (Ω ) ∩ Hm 0 (Ω ). We do this in the same way as in theorem A.77 (which in fact yields the case m = 1); that is, we consider the elliptic boundary value problem   (−∆)m u = f in Ω , ju (A.76) ∂  =0 on ∂ Ω , 0 ≤ j ≤ m − 1 . j ∂ν Since Am is an elliptic operator which is also uniformly strongly elliptic (see e.g. Dautray-Lions, [DL88, sct. VII.5]), given f ∈ H−m (Ω ), by Lax-Milgram’s theorem 2 there is a unique u ∈ Hm 0 (Ω ), solution of (A.76). If in addition f ∈ L (Ω ), then 2m u ∈ H (Ω ), as follows from theorem 5.3 of chapter 2, sct. 5.3, and remark 1.3 of chapter 2, sct. 1.4, of Lions-Magenes, [LM72] (see also Dautray-Lions, [DL88, sct. VII.1.7]). Consequently, we can consider the operator 1/2

Λ = Am = (−∆)m/2

A.6

Sobolev Spaces of Vector Valued Functions

361

2 as an operator that defines the interpolation spaces [Hm 0 (Ω ), L (Ω )]θ , in accord with (A.42). Recalling then (A.69), we have that m s 2 ˜ s (Ω ) , 2 m ) = dom((−∆)s/2 ) = H Hs0 (Ω ) = Hm 0 (Ω ), L (Ω ) 1−s/m = dom((−∆)

as per definition (A.75). In particular, for m = 1 we are reduced to the case X = H10 (Ω ), Y = L2 (Ω ), with A1 = −∆ and [X , Y]θ = dom((−∆)(1−θ )/2 ) . Taking θ = 0 yields then that dom((−∆)1/2 ) = X = H10 (Ω ) , and (A.41) reduces to the identity hu, viX = h∇u, ∇vi , which corresponds to the choice in H10 (Ω ) of the scalar product inducing the norm (A.72). In this way, the operator (−∆)1/2 can be formally related (but not identified) with the operator ∇. 5. We conclude this section by mentioning the following result on the operator −∆. THEOREM A.79 Let Ω be a bounded domain of RN . The operator −∆ generates an analytic semigroup in L2 (Ω ). PROOF This is a consequence of a more general result, concerning strongly elliptic operators (even with variable coefficients). See e.g. Pazy, [Paz83, sct. 7.2, thm. 2.7].

A.6 Sobolev Spaces of Vector Valued Functions A.6.1 Lebesgue and Sobolev Spaces In this section we introduce the Lebesgue and Sobolev spaces L p (a, b; X ) and where ]a, b[ ⊆ R is an interval, and X is a separable Hilbert space. For 1 ≤ p ≤ +∞, we define L p (a, b; X ) to be the space of the (equivalence classes of) functions u : ]a, b[→ X which are strongly measurable, and such that

Hm (a, b; X ),

Z b a

p ku(t)kX dt < +∞

(A.77)

362

Appendix: Selected Results from Analysis

if 1 ≤ p < +∞, or sup ess{ku(t)kX : a < t < b} < +∞ . In (A.77), the integral is meant in the sense of Bochner (see e.g. Yosida, [Yos80, sct. V.5]), It can be shown that L p (a, b; X ) is a Banach space, with respect to the norms kuk p :=

b

Z a

1/p

p

ku(t)kX dt

,

1 ≤ p < +∞ ,

kuk∞ := sup ess{ku(t)kX : a < t < b} . If p = 2, L2 (a, b; X ) is a Hilbert space, with respect to the scalar product hu, vi0 :=

Z b

hu(t), v(t)iX dt .

a

We can introduce in L p (a, b; X ) the notion of weak (i.e., distributional) derivatives with respect to t, in a manner totally analogous to definition A.63. For k ∈ N and u ∈ L p (a, b; X ), we denote by u(k) its strong derivative of order k. For m ∈ N we define then Hm (a, b; X ) := {u ∈ L2 (a, b; X ) : u( j) ∈ L2 (a, b; X ), 0 ≤ j ≤ m} . This is a Hilbert space, with respect to the scalar product m

hu, vim :=

∑

Z b

hu( j) (t), v( j) (t)iX dt .

j=0 a

A.6.2 The Intermediate Derivatives Theorem In the sequel, we consider two Hilbert spaces X and Y, with X ,→ Y, densely and continuous. For integer m ≥ 0, we set W(a, b; X , Y) := {u ∈ L2 (a, b; X ) : u(m) ∈ L2 (a, b; Y)} ; when there is no possibility of confusion, we abbreviate W(a, b; X , Y) = W. Then, recalling section A.2.10, we have the following result: THEOREM A.80 Let u ∈ W. Then, for 0 ≤ j ≤ m, u( j) ∈ L2 (a, b; [X , Y] j/m ) . Moreover, if 0 ≤ j ≤ m − 1 and θ j := m1 j + 12 , u( j) ∈ C([a, b]; [X , Y]θ j ) .

A.7

The Spaces H(div, Ω ) and H(curl, Ω )

363

The linear maps u 7→ u( j) of W into L2 (a, b; [X , Y] j/m ), 0 ≤ j ≤ m (respectively, into C([a, b]; [X , Y]θ j ), 0 ≤ j ≤ m − 1), are continuous. In particular, there are positive constants C1 and C2 , with C1 independent of the length of ]a, b[, such that for all u ∈ W and 1 ≤ j ≤ m − 1, 1−θ

θ

j max ku( j) (t)k[X ,Y ]θ ≤ C1 kukL2 (a,b; ku(m) kL2j (a,b;Y ) +C2 kukL2 (a,b;X ) . X)

a≤t ≤b

j

PROOF See e.g. Lions-Magenes, [LM72, scts. 1.2, 1.3]. As a consequence, recalling the definition (A.67) of Hs (Ω ), as well as theorem A.75, we deduce from theorem A.80 the following imbeddings results. THEOREM A.81 Let m ≥ r ≥ −1, and 0 ≤ k ≤ m. Assume that u ∈ {u ∈ L2 (a, b; Hm (Ω )) : u(k) ∈ L2 (a, b; Hr (Ω ))} . −r Let p := m − kj (m − r) and q := p − m2k . Then, for 0 ≤ j ≤ k − 1,

u( j) ∈ L2 (a, b; H p (Ω )) ∩ C([a, b]; Hq (Ω )) . Moreover, there exist positive constants C1 and C2 , with C1 independent of the length of ]a, b[, such that 1−( j+1/2)/k

( j+1/2)/k

max ku( j) (t)kq ≤ C1 kukL2 (a,b;Hm (Ω )) ku(k) kL2 (a,b;Hr (Ω )) +C2 kukL2 (a,b;Hm (Ω )) .

a≤t ≤b

We conclude with the following compact imbedding result. THEOREM A.82 Let X , Y and Z be reflexive Banach spaces, with X ,→ Y ,→ Z, the injection X ,→ Y being compact. Let p, q ∈ ]1, +∞[. Then the injection {u ∈ L p (a, b; X ) : u0 ∈ Lq (a, b; Z)} ,→ L p (a, b; Y) , which is continuous, is also compact. PROOF See e.g. Lions, [Lio69, sct. 1.5].

A.7 The Spaces H(div, Ω ) and H(curl, Ω ) In this section we consider certain subspaces of the space L2 (Ω ), which are particularly suited to the mathematical study of various systems of PDEs which arise

364

Appendix: Selected Results from Analysis

in fluid and electromagnetic dynamics. For a proof of the results we mention without justification, we refer e.g. to Duvaut-Lions, [DL69, sct. 5.7], or to Foias-Temam, [FT78].

A.7.1 Notations We assume that Ω is a bounded open domain of RN , simply connected, and that its boundary ∂ Ω is at least of Lipschitz class; we denote by ~ν the outward unit normal to ∂ Ω . As usual, we denote by k · k and h · , · i the norm and scalar product in L2 (Ω ). In addition, we adopt the following CONVENTION A.83 If ~u = (u1 , . . . , uN ) : Ω → RN , N ≥ 2, is a vector valued function, and X is a space of scalar functions on Ω , such as L2 (Ω ) or C(Ω ), then, with abuse of notation, we write ~u ∈ X to mean that all the components of ~u are in X . In particular, if the context is clear and there is no danger of confusion, we use the same notation X to denote the space X N as well. We consider the differential operators “div” and “curl”, formally defined as follows. Given ~u = (u1 , . . . , uN ) : RN → RN , for j = 1, . . . , N we set ∂ j = ∂ /∂ x j , and N

div~u :=

∑ ∂ ju j ;

j=0

thus, x 7→ (div~u)(x) is a scalar function. Then, if N = 3 we define curl~u := (∂2 u3 − ∂3 u2 , ∂3 u1 − ∂1 u3 , ∂1 u2 − ∂2 u1 ) , while if N = 2 we define curl~u := ∂1 u2 − ∂2 u1 ; thus, x 7→ (curl~u)(x) is a scalar function if N = 2, and a vector function if N = 3.

A.7.2 The Space H(div, Ω ) 1.

We set H(div, Ω ) := {~u ∈ L2 (Ω ) : div~u ∈ L2 (Ω )} .

This linear subspace of L2 (Ω ) is a Hilbert space with respect to the norm defined by k~uk2div := k~uk2 + k div~uk2 . Clearly, H1 (Ω ) ,→ H(div, Ω ); in fact, we have THEOREM A.84 The space C∞ (Ω ) is dense in H(div, Ω ).

The Spaces H(div, Ω ) and H(curl, Ω )

A.7

365

As a consequence, the NORMAL COMPONENT trace operator ~u 7→ ~ν ·~u can be extended by continuity from a linear continuous map of C∞ (Ω ) into C∞ (∂ Ω ), to a linear continuous map, still denoted by ~u 7→ ~ν ·~u, of H(div, Ω ) into H−1/2 (∂ Ω ). More precisely, denoting by h · , · i1/2 the duality pairing between H−1/2 (∂ Ω ) and H1/2 (∂ Ω ), the trace ~ν ·~u is defined as follows. Given any ψ ∈ H1/2 (∂ Ω ), we choose Ψ ∈ H1 (Ω ) such that γ0 (Ψ ) = ψ (this is possible by the surjectivity of γ0 ; see theorem A.73). Then, we define h~ν ·~u, ψi1/2 := hdiv~u,Ψ i + h~u, ∇Ψ i .

(A.78)

Indeed, we easily see that the right side of (A.78) is independent of the choice of Ψ (as long as γ0 (Ψ ) = ψ, of course), and that, if ~u ∈ H(div, Ω ), it is continuous in Ψ with respect to the H1 norm. Since the dependence of Ψ on ψ is also continuous, the right side of (A.78) depends continuously on ψ; hence, it defines an element of H−1/2 (∂ Ω ). This element is precisely the desired normal component of ~u. Note that (A.78) generalizes the well known integration by parts formula Z

(~ν ·~u)v ds =

Z

Z

(div~u)v ds + Ω

∂Ω

~u · ∇v ds ,

Ω

which is valid for all ~u and v in the space C1 (Ω ), which is dense in H(div, Ω ). 2.

We can then introduce the following subspaces of H(div, Ω ). H0 (div, Ω ) := {~u ∈ H(div, Ω ) : ~ν ·~u = 0} , H0 (div, Ω ) := {~u ∈ H(div, Ω ) : div~u = 0} , H00 (div, Ω ) := H0 (div, Ω ) ∩ H0 (div, Ω ) .

THEOREM A.85 The space u = 0} C∞ u ∈ C∞ 0 (Ω ) : div~ 0,div (Ω ) := {~ is dense in H00 (div, Ω ). Moreover, setting H1ν (Ω ) := {~u ∈ H1 (Ω ) : ~ν ·~u = 0} ,

(A.79)

the F RIEDRICHS ’ INEQUALITY k~uk ≤ C (k div~uk + k curl~uk) holds for all ~u ∈ H1ν (Ω ), with C independent of ~u.

A.7.3 The Space H(curl, Ω ) 1.

We set H(curl, Ω ) := {~u ∈ L2 (Ω ) : curl~u ∈ L2 (Ω )} .

(A.80)

366

Appendix: Selected Results from Analysis

This linear subspace of L2 (Ω ) is a Hilbert space with respect to the norm defined by k~uk2curl := k~uk2 + k curl~uk2 . Clearly, H1 (Ω ) ,→ H(curl, Ω ); in fact, we have THEOREM A.86 The space C∞ (Ω ) is dense in H(curl, Ω ). As a consequence, the TANGENTIAL COMPONENT trace operator ~u 7→ ~ν ×~u can be extended by continuity from a linear continuous map of C∞ (Ω ) into C∞ (∂ Ω ), to a linear continuous map, still denoted by ~u 7→ ~ν ×~u, of H(curl, Ω ) into H−1/2 (∂ Ω ). This is done exactly as for the definition of the normal component ~ν ·~u; that is, we define ~ν ×~u ∈ H−1/2 (∂ Ω ) by the identity

h~ν ×~u, ψi1/2 := hcurl~u,Ψ i − h~u, curlΨ i ,

(A.81)

for ~u ∈ H(curl, Ω ), ψ ∈ H1/2 (∂ Ω ), and Ψ ∈ H1 (Ω ) such that γ0 (Ψ ) = ψ. Note that (A.81) generalizes the well known integration by parts formula Z

~ν ×~u ·~v ds =

∂Ω

Z

curl~u ×~v ds −

Ω

Z

~u × curl~v ds ,

Ω

which is valid for all ~u and ~v in the space C1 (Ω ), which is dense in H(curl, Ω ). 2.

We can then introduce the following subspaces of H(curl, Ω ). H0 (curl, Ω ) := {~u ∈ H(curl, Ω ) : ~ν ×~u = ~0} , H0 (curl, Ω ) := {~u ∈ H(curl, Ω ) : curl~u = ~0} , H00 (curl, Ω ) := H0 (curl, Ω ) ∩ H0 (curl, Ω ) .

THEOREM A.87 The space u = ~0} C∞ u ∈ C∞ 0 (Ω ) : curl~ 0,curl (Ω ) := {~ is dense in H00 (curl, Ω ). Moreover, setting H1τ (Ω ) := {~u ∈ H1 (Ω ) : ~ν ×~u = ~0} , the Friedrichs’ inequality (A.80) also holds for all ~u ∈ H1τ (Ω ).

(A.82)

A.7

The Spaces H(div, Ω ) and H(curl, Ω )

367

A.7.4 Relations between H(div, Ω ) and H(curl, Ω ) In this section we report some results on spaces that are constructed from both H(div, Ω ) and H(curl, Ω ). In particular, we present a well known decomposition theorem of L2 (Ω ). THEOREM A.88 Let H1ν (Ω ) and H1τ (Ω ) be the spaces defined in (A.79) and (A.82). Then, H0 (curl, Ω ) ∩ H(div, Ω ) = H1τ (Ω ) , H(curl, Ω ) ∩ H0 (div, Ω ) = H1ν (Ω ) , and the norm defined by k~uk2X := k curl~uk2 + k div~uk2 , where X denotes either of the spaces H1ν (Ω ) or H1τ (Ω ), is equivalent to the H1 norm in X . Consequently, if ~u ∈ X is such that div~u = 0 and curl~u = ~0, then ~u = ~0. PROPOSITION A.89 For all ~a ∈ H0 (curl, Ω ), curl~a ∈ H00 (div, Ω ). PROOF Since div(curl~a) = ~0, curl~a ∈ H0 (div, Ω ), so that its normal component is defined, with ~ν · curl~a ∈ H−1/2 (∂ Ω ). Given then arbitrary ϕ ∈ H1/2 (∂ Ω ), let Φ ∈ H1 (Ω ) be such that γ0 (Φ) = ϕ. By density, we can without loss of generality assume that Φ ∈ H2 (Ω ), so that γ0 (∇Φ) ∈ H1/2 (∂ Ω ). Then, recalling (A.78) and (A.81), h~ν · curl~a, ϕi1/2 = hcurl~a, ∇Φi = h~ν ×~a, γ0 (∇Φ)i1/2 = 0 .

(A.83)

PROPOSITION A.90 For all ψ ∈ H10 (Ω ), ∇ψ ∈ H00 (curl, Ω ). PROOF Since curl(∇ψ) = ~0, ∇ψ ∈ H0 (curl, Ω ), so that its tangential component is defined, with ~ν × ∇ψ ∈ H−1/2 (∂ Ω ). Given then arbitrary ϕ ∈ H1/2 (∂ Ω ), let Φ ∈ H2 (Ω ) be such that γ0 (Φ) = ϕ. Then, recalling (A.81) and (A.78), h~ν × ∇ψ, ϕi1/2 = −h∇ψ, curl Φi = h~ν · curl Φ, ψi1/2 = 0 .

(A.84)

368

Appendix: Selected Results from Analysis

We now set V0 := H0 (curl, Ω ) ∩ H0 (div, Ω ) , and note that, by Friedrichs’ inequality (A.80), the norm defined by k~ukV0 := k curl~uk is a norm in V0 equivalent to the one induced by H1 (Ω ). PROPOSITION A.91 For all ~f ∈ H00 (div, Ω ), there exists a unique ~u ∈ V0 such that curl~u = ~f . PROOF Define a bilinear form a on V0 × V0 by a(u, v) := hcurl~u, curl~vi . Then, a is bilinear and coercive on V0 ; indeed, for all ~u ∈ V0 , a(u, u) = k curl~uk2 = k~uk2V0 . Let f ∈ V00 be defined by h f ,~viV 0 ×V0 := h~f , curl~vi , 0

~v ∈ V0 .

By the Lax-Milgram theorem A.35, there is a unique ~u ∈ V0 , such that for all ~v ∈ V0 , hcurl~u, curl~vi = h~f , curl~vi .

(A.85)

We now show that ~w := curl~u − ~f = ~0, as desired. Indeed, we can compute curl~w at least in distributional sense; more precisely, since ~w ∈ L2 (Ω ), recalling (A.61) and (A.60) we have that, for all ~ϕ ∈ D(Ω ), hcurl~w, ~ϕ iD = h~w, curl ~ϕ iD = h~w, curl ~ϕ i .

(A.86)

Given ~ϕ ∈ D(Ω ), let β ∈ C∞ (Ω ) ∩ C0 (Ω ) be defined as the solution to the elliptic BVP ( −∆β = div ~ϕ , β |∂ Ω = 0 . Then, by proposition A.90, ~ϕ + ∇β ∈ V0 ; therefore, by (A.85), h~w, curl ~ϕ i = h~w, curl(~ϕ + ∇β )i = 0 .

(A.87)

Together with (A.86), (A.87) implies that curl~w = ~0 in D0 (Ω ). But then, this also means that curl~w ∈ L2 (Ω ), so that ~w ∈ H0 (curl, Ω ). Next, we note that div~w = div ~f = 0, so that ~w ∈ H0 (div, Ω ) as well. In particular, the normal component ~ν · ~w

A.7

The Spaces H(div, Ω ) and H(curl, Ω )

369

is defined in H−1/2 (∂ Ω ); since ~ν ×~u = ~0 and ~ν · ~f = 0, recalling (A.83) we deduce that ~ν · ~w = 0. In conclusion, we have that ~w ∈ H1ν (Ω ), and curl~w =~0, div~w = 0. By the last part of theorem A.88, we conclude that ~w = ~0, as claimed. Finally, to prove the uniqueness of ~u it is sufficient to note that if ~v ∈ V0 is also such that curl~v = ~f , then the difference~z := ~u −~v is such that curl~z = ~0 ,

div~z = 0 ,

~ν ×~z = ~0 .

Hence, arguing as we did above for ~w, we deduce that~z = ~0. REMARK A.92 We can also determine ~u by resorting to general results in the classical theory of elliptic systems (see e.g. Agmon-Douglis-Nirenberg, [ADN64]). Indeed, it is sufficient to note that ~u can be determined as the solution of the system    curl~u + ∇p = ~f , (A.88) div~u = 0,  ~ ν ×~u = ~0 . In fact, it turns out that this system is strongly elliptic, that its boundary conditions are complementing, and that the source data {~f , 0} are orthogonal to the range of the adjoint of the operator L := (curl +∇, div) (which is L∗ = (curl −∇, − div)). Hence, (A.88) has a unique solution (~u, p). However, ∇p = ~0, because p is also a solution of the elliptic Neumann BVP ( −∆p = − div ~f = 0 , ∂p = ~ν · ∇p = ~ν · (~f − curl~u) = 0 ∂~ν (the last step follows from (A.83)). Hence, curl~u = ~f . In fact, the converse of proposition A.91 is also true. THEOREM A.93 The following identities hold, as sets and as Hilbert spaces: H00 (div, Ω ) = {~u ∈ L2 (Ω ) : (∃~a ∈ H(curl, Ω ) : ~u = curl~a)} , H00 (curl, Ω ) = {~u

2

∈ L (Ω ) : (∃ ψ

∈ H10 (Ω ) :

~u = −∇ψ)} .

(A.89) (A.90)

PROOF The inclusions ⊇ are a consequence of propositions A.89 and A.90. For the reverse inclusion in (A.89), use proposition A.91, with ~f = ~u. For the inclusion ⊆ in (A.90), given ~u ∈ H00 (curl, Ω ) define ψ as the solution of the elliptic Dirichlet BVP ( −∆ψ = div~u , ψ| ∂ Ω = 0 .

370

Appendix: Selected Results from Analysis

Since div~u ∈ H−1 (Ω ), by theorem A.77 we deduce that ψ is uniquely determined in H10 (Ω ). Let~z := ~u + ∇ψ. Then,    curl~z = curl~u = ~0 , div~z = div~u + ∆ψ = 0 ,  ~ ν ×~z = ~ν × ∇ψ = ~0 (the last step following as in (A.84)). Hence,~z = ~0, by the last part of theorem A.88; that is, ~u = −∇ψ, as desired. We can then finally state the following orthogonality result. THEOREM A.94 Let Ω ⊂ R3 be a bounded, simply connected open domain, with a Lipschitz boundary ∂ Ω . Then, L2 (Ω ) = H00 (div, Ω ) ⊕ H0 (curl, Ω ) . In other words, for all ~u ∈ L2 (Ω ), there exist ~a ∈ H0 (curl, Ω ) and ψ ∈ H1 (Ω ), such that ~u = curl~a − ∇ψ .

(A.91)

The functions ψ and ~a of theorem A.93 are usually called the SCALAR and VECTOR POTENTIALS of ~u; in general, they are not uniquely determined. PROOF The inclusion ⊇ is obvious. For the converse, we introduce the space V1 := H1 (Ω ) R of the equivalence classes Φ = [ϕ] of functions in H1 (Ω ), with respect to the equivalence relation defined by f ∼g

def.

⇐⇒

∃ c ∈ R : f (x) − g(x) = c a.e.

We denote such equivalence classes by capital letters like Ψ , and any of their representatives by a lower case letter like ψ. The space V1 is a Hilbert space, with respect to the norm kΨ kV1 = k∇ψk ,

ψ ∈Ψ

(A.92)

(see e.g. Neˇcas, [Ne£67, sct. 2.7.2]; note that the right side of (A.92) is independent of the particular representative ψ of Ψ ). Consider then the bilinear form on V1 defined by a(Φ,Ψ ) := h∇ϕ, ∇ψi , ϕ ∈ Φ , ψ ∈Ψ .

A.8

Almost Periodic Functions

371

This form is continuous and coercive on V1 ; in particular, a(Φ, Φ) = k∇ϕk2 = kΦk2V1 . The map Φ 3 Φ 7→ −h~u, ∇ϕi ,

ϕ ∈Φ,

is linear continuous on V1 ; therefore, by the Lax-Milgram theorem A.35, there is a unique Ψ ∈ V1 such that for all Φ ∈ V1 , a(Ψ , Φ) = h∇ψ, ∇ϕi = h~u, ∇ϕi .

(A.93)

Fix now any ψ ∈ Ψ (thus, ψ ∈ H1 (Ω )). Given any ζ ∈ H10 (Ω ), we compute that, because of (A.93), h− div(∇ψ +~u), ζ iH−1 (Ω )×H1 (Ω ) = h∇ψ +~u, ∇ζ i = 0 . 0

Consequently, ∇ψ +~u ∈ H0 (div, Ω ); therefore, its normal component is defined in H−1/2 (∂ Ω ). Given any α ∈ H1/2 (∂ Ω ), let A ∈ H1 (Ω ) be such that γ0 (A) = α. Then, again by (A.93), h~ν · (∇ψ +~u), αi1/2 = (∇ψ +~u, ∇A) = 0 . Consequently, ∇ψ +~u ∈ H00 (div, Ω ), so that, by (A.89) of theorem A.93, there is ~a ∈ H0 (curl, Ω ) such that ∇ψ +~u = curl~a, as desired in (A.91).

A.8 Almost Periodic Functions In this section we briefly recall some basic facts concerning almost periodic functions, valued in a separable Banach space X . All the material in this section is taken from Amerio-Prouse, [AP71]. DEFINITION A.95 A subset J ⊆ R is said to be RELATIVELY DENSE, if there is ` > 0 such that any interval of R of length ` contains at least one point of J . DEFINITION A.96 Let f : R → X be continuous and, for τ ∈ R, define the translated function t 7→ fτ (t) := f (t + τ). The function f is said to be ALMOST PERIODIC if for all ε > 0 there is a relatively dense set T ⊆ R such that for all τ ∈ T , sup k f (t + τ) − f (t)kX < ε . t ∈R

Each τ ∈ T is called an ε-ALMOST PERIOD of f .

372

Appendix: Selected Results from Analysis

The following characterization of almost periodic functions is generally known as B OCHNER ’ S THEOREM . THEOREM A.97 The following conditions are equivalent: 1. The function f : R → X is almost periodic; 2. The set of the translations ( fτ )τ ∈R is relatively compact in the space Cb (R), endowed with the sup norm; 3. There exists a relatively dense sequence (τm )m∈N such that the sequence of translated functions ( f (· + τm ))m∈N is relatively compact in Cb (R). PROOF See e.g. Amerio-Prouse, [AP71, ch. 1]. In particular, since for all x ∈ X and λ ∈ R, the function t 7→ eiλt x is periodic, we deduce that all trigonometric polynomials of the form N

∑ eiλkt ak , k=1

with λk ∈ R and ak ∈ X , are almost periodic. The following theorem characterizes almost periodic functions as the sum of a uniformly convergent trigonometric series. THEOREM A.98 A function f : R → RN is almost periodic if and only if there are sequences (ak )k∈N ⊆ RN and (λk )k∈N ⊂ R such that for all t ∈ R, ∞

f (t) =

∑ eiλkt ak . k=1

PROOF See e.g. Amerio-Prouse,[AP71, ch. 2].

Bibliography

[Ada78] R.A. Adams. Sobolev spaces. Academic Press, Inc., New York, 1978. [ADN64] S. Agmon, A. Douglis, and L. Nirenberg. Estimates near the boundary for the solution of elliptic differential equations satisfying general boundary values, II. Comm. Pure Appl. Math., 17:35–92, 1964. [Ama90] H. Amann. Ordinary Differential Equations: An Introduction to Nonlinear Analysis. De Gruyter, Berlin, 1990. [AP71] L. Amerio and G. Prouse. Almost-periodic functions and functional equations. (The University Series in Higher Mathematics.). Van Nostrand Reinhold Company VIII, New York, 1971. [AP93] A. Ambrosetti and G. Prodi. A Primer of Nonlinear Analysis. Cambridge University Press, Cambridge, 1993. [ASY96] K.T. Alligood, T.D. Sauer, and J.A. Yorke. Chaos. An Introduction to Dynamical Systems. Springer-Verlag, New York, 1996. [Bal73] J.M. Ball. Stability theory for an extensible beam. J. Diff. Eqs., 14:399–418, 1973. [BdMCR98] L. Boutet de Monvel, I.D. Chueshov, and A.V. Rezounenko. Inertial manifolds for retarded semilinear parabolic equations. Nonlinear Analysis, 34:907–925, 1998. [BF95] A. Bensoussan and F. Flandoli. Stochastic inertial manifold. Stochastics Rep., 53(1-2):13–39, 1995. [BJ89] P.W. Bates and C.K.R.T. Jones. Invariant manifolds for semilinear partial differential equations. In U. Kirchgrabner and H.O. Walther, editors, Dynamics Reported, Vol. 2, pages 1–38. John Wiley & Sons, New York, 1989. [BL76] J. Bergh and J. Löfström. Interpolation Spaces. An Introduction, volume 223 of Grundlehren der mathematischen Wissenschaften. Springer-Verlag, Berlin, 1976. [Bre73] H. Brezis. Opérateurs Maximaux Monotones. North-Holland, Amsterdam, 1973. [BV92] A. V. Babin and M. I. Vishik. Attractors of Evolution Equations. North Holland, Amsterdam, 1992.

373

374

Bibliography [CF88] P. Constantin and C. Foias. Navier-Stokes Equations. The University of Chicago Press, Chicago, 1988.

[CFNT86] P. Constantin, C. Foias, B. Nicolaenko, and R. Temam. Nouveaux resultats sur les variétés inertielles pour les équations différentielles dissipatives. C. R. Acad. Sci., Paris, Ser. I, 302:375–378, 1986. [CFNT89] P. Constantin, C. Foias, B. Nicolaenko, and R. Temam. Integral Manifolds and Inertial Manifolds for Dissipative Partial Differential Equations, volume 70 of Applied Mathematical Sciences. Springer-Verlag, Berlin, 1989. [Chu95] I.D. Chueshov. Approximate inertial manifolds of exponential order for semilinear parabolic equations subjected to additive white noise. J. Dyn. Differ. Equations, 7(4):549–566, 1995. [CI74] N. Chaffee and E. Infante. A bifurcation problem for a nonlinear parabolic equation. Appl. Anal., 4:17–37, 1974. [CL55] E.A. Coddington and N. Levinson. Theory of Ordinary Differential Equations. McGill-Hill Book Company, Inc., 1955. [CL99] T. Caraballo and J.A. Langa. Tracking properties of trajectories on random attracting sets. Stochastic Anal. Appl., 17(3):339–358, 1999. [CS01] I.D. Chueshov and M. Scheutzow. Inertial manifolds and forms for stochastically perturbed retarded semilinear parabolic equations. J. Dyn. Differ. Equations, 13(2):355–380, 2001. [DG91] F. Demengel and J.M. Ghidaglia. Some remarks on the smoothness of inertial manifolds. Nonlinear Anal., Theory Methods Appl., 16(1):79– 87, 1991. [DL69] G. Duvaut and J.L. Lions. Les Inéquations en Mécanique et en Physique. Dunod-Gauthier-Villars, Paris, 1969. [DL88] R. Dautray and J.-L. Lions. Mathematical Analysis and Numerical Methods for Science and Technology. Volume 2: Functional and Variational Methods. Springer-Verlag, Berlin, 1988. [DL90] R. Dautray and J.-L. Lions. Mathematical Analysis and Numerical Methods for Sciences and Technology. Volume 3: Spectral Theory and Applications. Springer-Verlag, Berlin, 1990. [DLS03] J. Duan, K. Lu, and B. Schmalfuss. Invariant manifolds for stochastic partial differential equations. Ann. Probab., 31(4):2109–2135, 2003. [EFK98] A. Eden, C. Foias, and V. Kalantarov. A remark on two constructions of exponential attractors for α-contractions. J. Dyn. Differ. Equations, 10(1):37–45, 1998.

Bibliography

375

[EFNT94] A. Eden, C. Foias, B. Nicolaenko, and R. Temam. Exponential Attractors for Dissipative Evolution Equations. Research in Applied Mathematics. Wiley and Masson, 1994. [EM93] A. Eden and A. Milani. Exponential attractors for extensible beam equations. Nonlinearity, 6:457–479, 1993. [EM95] A. Eden and A. Milani. On the convergence of attractors and exponential attractors for singularly perturbed hyperbolic equations. Turkish Math. J., 19:102–117, 1995. [EMN92] A. Eden, A. Milani, and B. Nicolaenko. Finite dimensional exponential attractors for semilinear wave equations with damping. J. Math. An. Appl., 169(2):408–419, 1992. [EN00] K.-J. Engel and R. Nagel. One-Parameter Semigroups for Linear Evolution Equations., volume 194 of Graduate Texts in Mathematics. Springer-Verlag, Berlin, 2000. [Fal85] K.J. Falconer. The Geometry of Fractal Sets, volume 85 of Cambridge Tracts in Mathematics. Cambridge University Press, Cambridge, 1985. [Fal90] K. Falconer. Fractal Geometry: Mathematical Foundations and Applications. John Wiley & Sons, Chichester, 1990. [Fei78] M. J. Feigenbaum. Qualitative universality for a class of nonlinear transformations. J. Stat. Phys., 19(1):25–52, 1978. [FGMZ03] P. Fabrie, C. Galusinski, A. Miranville, and S. Zelik. Uniform Exponential Attractors for a Singularly Perturbed Damped Wave Equation. Preprint, 2003. [FNST85] C. Foias, B. Nicolaenko, G.R. Sell, and R. Temam. Variétés inertielles pour l’équation de Kuramoto-Sivashinsky. C. R. Acad. Sci., Paris, Ser. I, 301:285–288, 1985. [FNST88] C. Foias, B. Nicolaenko, G.R. Sell, and R. Temam. Inertial manifolds for the Kuramoto-Sivashinsky equation and an estimate of their lowest dimension. J. Math. Pures Appl., IX., 67(3):197–226, 1988. [FST85] C. Foias, G.R. Sell, and R. Temam. Variétés inertielles des équations différentielles dissipatives. (Inertial manifolds for dissipative differential equations). C. R. Acad. Sci., Paris, Ser. I, 301:139–141, 1985. [FST88] C. Foias, G.R. Sell, and R. Temam. Inertial manifolds for nonlinear evolutionary equations. J. Differ. Equations, 73:309–353, 1988. [FST89] C. Foias, G.R. Sell, and E.S. Titi. Exponential tracking and approximation of inertial manifolds for dissipative nonlinear equations. J. Dyn. Differ. Equations, 1(2):199–244, 1989.

376

Bibliography [FT78] C. Foias and R. Temam. Remarques sur les équations de Navier-Stokes stationnaires et les phénomènes successifs de bifurcation. Ann. Sc. Norm. Sup. Pisa, Ser. IV, 5:29–63, 1978.

[GGMP03] S. Gatti, M. Grasselli, A. Miranville, and V. Pata. A Construction of a Robust Family of Exponential Attractors. Preprint, 2003. [GH83] J. Guckenheimer and P. Holmes. Nonlinear Oscillations, Dynamical Systems, and Bifurcations of Vector Fields. Springer-Verlag, New York, 1983. [GP02] M. Grasselli and V. Pata. On the damped semilinear wave equation with critical exponent. In Proc. of IV Int. Conf. on Dyn. Systs. and PDEs, May 24-27, Wilmington, NC, USA, 2002. [GT83] D. Gilbarg and N.S. Trudinger. Elliptic Partial Differential Equations of Second Order, volume 224 of Grundlehren der mathematischen Wissenschaften. Springer-Verlag, Berlin, 1983. [GV97] A.Y. Goritskij and M.I. Vishik. Local integral manifolds for a nonautonomous parabolic equation. J. Math. Sci., New York, 85(6):2428– 2439, 1997. [Had01] J. Hadamard. Sur l’iteration et les solutions asymptotiques des equations différentielles. Bull. Soc. Math. France, 29:224–228, 1901. [Hal88] J. Hale. Asymptotic Behavior of Dissipative Systems, volume 25 of AMS Math. Surveys and Monographs. Providence, Rhode Island, 1988. [Har91] A. Haraux. Systemes Dynamiques Dissipatifs et Applications. Masson, Paris, 1991. [Hen81] D. Henry. Geometric Theory of Semilinear Parabolic Equations, volume 850 of Lecture Notes in Mathematics. Springer-Verlag, 1981. [Hen85] D. Henry. Some infinite-dimensional Morse-Smale systems defined by parabolic partial differential equations. J. Differ. Equations, 59:165– 205, 1985. [HP76] M. Hénon and Y. Pomeau. Two strange attractors with a simple structure. In R. Temam, editor, Turbulence and Navier-Stokes Equations, volume 565 of Lecture Notes Math., pages 29–68. Springer-Verlag, Berlin, 1976. [HR88] J. K. Hale and G. Raugel. Upper semicontinuity of the attractor for a singularly perturbed hyperbolic equation. J. Diff. Eqs., 73(2):197–214, 1988. [HR90] J. K. Hale and G. Raugel. Lower semicontinuity of the attractor for a singularly perturbed hyperbolic equation. J. Dyn. Diff. Eqs., 2(1):19– 67, 1990.

Bibliography

377

[HS93] M. Hirsch and S. Smale. Differential Equations, Dynamical Systems, and Linear Algebra. Academic Press, Boston, 1993. [Jol89] M.S. Jolly. Explicit construction of an inertial manifold for a reaction diffusion equation. J. Differ. Equations, 78(2):220–261, 1989. [JS87] D. W. Jordan and P. Smith. Nonlinear Ordinary Differential Equations. Clarendon Press, Oxford, 1987. [JT96] D.A. Jones and E.S. Titi. C1 approximations of inertial manifolds for dissipative nonlinear equations. J. Differ. Equations, 127(1):54–86, 1996. [Kat95] T. Kato. Perturbation Theory for Linear Operators. Classics in Mathematics. Springer-Verlag, Berlin, 1995. [Kom70] Y. Komura. Nonlinear semigroups in Hilbert spaces. In Proc. Int. Conf. Funct. Anal. Rel. Topics, Tokyo 1969, pages 260–268. 1970. [KS02a] N. Koksch and S. Siegmund. Inertial manifolds for nonautonomous dynamical systems and for nonautonomous dynamical systems. CD ROM Proceedings of the Equadiff 10, Prague 2001, pages 221–266, 2002. [KS02b] N. Koksch and S. Siegmund. Pullback attracting inertial manifolds for nonautonomous dynamical systems. J. Dyn. Differ. Equations, 14(4):889–941, 2002. [KS03] N. Koksch and S. Siegmund. Cone invariance and squeezing properties for inertial manifolds of nonautonomous evolution equations. Banach Center Publications, vol. 60, Warzawa, pages 27–48, 2003. [Kur66] C. Kuratowski. Topology, volume I. Academic Press, New York, 1966. [Lio69] J. L. Lions. Quelques Méthodes de Résolution des Problémes aux Limites non Linéaires. Dunod, Paris, 1969. [Lio73] J. L. Lions. Perturbations Singulières dans les Problémes aux Limites et en Contrôle Optimal, volume 323 of Lect. Notes Math. SpringerVerlag, Berlin, 1973. [LL99] Y. Latushkin and B. Layton. The optimal gap condition for invariant manifolds. Discrete and Continuous Dynamical Systems, 5(2):233– 268, 1999. [LM72] J. L. Lions and E. Magenes. Non-Homogeneous Boundary Value Problems, volume I. Springer-Verlag, New York, 1972. [Lor63] E.N. Lorenz. Deterministic nonperiodic flow. J. Atmospherical Sciences, 20(2):130–141, 1963.

378

Bibliography [LS89] M. Luskin and G.R. Sell. Approximation theories for inertial manifolds. RAIRO, Modelisation Math. Anal. Numer., 23(3):445–461, 1989. [Lya47] A.M. Lyapunov. Problème Gènèral de la Stabilité de Mouvement, volume 17 of Annals of Math. Studies. Princeton University Press, Princeton, 1947. [Lya92] A.M. Lyapunov. The General Problem of Stability of Motion. With a biography of Lyapunov by V.I. Smirnov and a bibliography of Lyapunov’s work by J.F. Barret. Transl. and ed. by A.T. Fuller. Taylor & Francis, London, 1992. [Mañ81] B. Mañé. On the Dimension of the Compact Invariant Sets of Certain Nonlinear Maps, volume 898 of Lecture Notes Math., pages 230–242. Springer-Verlag, Berlin, 1981. [MM76] J. Marsden and M. McCracken. Hopf Bifurcation and its Applications. Springer-Verlag, New York, 1976. [Moo92] F. C. Moon. Chaotic and Fractal Dynamics. Wiley, New York, 1992. [Mor87] X. Mora. Finite-dimensional attracting invariant manifolds for damped semilinear wave equations. In Contributions to Nonlinear Partial Differential Equations, Vol. II, Proc. 2nd Franco-Span. Colloq., Paris 1985, Pitman Res. Notes Math. Ser. 155, pages 172–183. 1987.

[MPS88] J. Mallet-Paret and G.R. Sell. Inertial manifolds for reaction diffusion equations in higher space dimensions. J. Am. Math. Soc., 1(4):804– 866, 1988. [MSM87] X. Mora and J. Solà-Morales. Existence and Nonexistence of Finite Dimensional Globally Attracting Invariant Manifolds in Semilinear Damped Wave Equations, pages 187–210. Springer-Verlag, New York, 1987. [Neˇc67] J. Neˇcas. Les Méthodes Directes en Théorie des Équations Elliptiques. Academia, Prague, 1967. [Nin93] H. Ninomiya. Inertial manifolds and squeezing properties. Lect. Notes Numer. Appl. Anal., 12:175–202, 1993. [Ole96] O. Oleinik. Some Asymptotic Problems in the Theory of Partial Differential Equations. Lezioni Fermiane. Cambridge Univ. Press, Cambridge, 1996. [Paz83] A. Pazy. Semigroups of Linear Operators and Applications to Partial Differential Equations. Springer-Verlag, New York, 1983. [Pee66] J. Peetre. Espaces d’interpolation et théorème de Sobolev. Ann. Inst. Fourier, 16:279 – 317, 1966.

Bibliography

379

[Per28] O. Perron. Über Stabilität und asymptotisches Verhalten der Integrale von Differentialgleichungssystemen. Math. Zeitschrift, 29:129–160, 1928. [Per29] O. Perron. Über Stabilität und asymptotisches Verhalten der Lösungen eines Systems endlicher Differenzengleichungen. J. Reine Angew. Math., 161:41–64, 1929. [Per30] O. Perron. Die Stabilitätsfrage bei Differentialgleichungen. Math. Zeitschrift, 32:703–728, 1930. [Per91] L. Perko. Differential Equations and Dynamical Systems. Texts in Applied Mathematics 7. Springer-Verlag, Berlin, 1991. [Rac92] R. Racke. Lectures on Nonlinear Evolution Equations. Vieweg, Braunschweig, 1992. [Rhe69] W.C. Rheinboldt. Local mapping relations and global implicit function theorems. Trans. Am. Math. Soc., 138:183–198, 1969. [Rob93] J.C. Robinson. Inertial manifolds and the cone condition. Dyn. Syst. Appl., 2(3):311–330, 1993. [Rob96] J.C. Robinson. The asymptotic completeness of inertial manifolds. Nonlinearity, 9:1325–1340, 1996. [Rob99] J.C. Robinson. Global attractors: Topology and finite-dimensional dynamics. J. Dyn. Differ. Equations, 11(3):557–581, 1999. [Rob01] J.C. Robinson. Infinite-Dimensional Dynamical Systems. An Introduction to Dissipative Parabolic PDEs and the Theory of Global Attractors. Cambridge Texts in Applied Mathematics. Cambridge University Press, Cambridge, 2001. [Rud73] W. Rudin. Functional Analysis. McGraw-Hill Series in Higher Mathematics. McGraw-Hill Book Comp., New York, 1973. [Rud74] W. Rudin. Real and Complex Analysis. 2nd ed. McGraw-Hill Series in Higher Mathematics. McGraw-Hill Book Comp., New York, 1974. [Spa82] C. Sparrow. The Lorenz Equations: Bifurcations, Chaos and Strange Attractors, volume 41 of Lecture Notes in Applied Mathematics. Springer-Verlag, Berlin, 1982. [SY02] G.R. Sell and Y. You. Dynamics of Evolutionary Equations, volume 143 of Applied Mathematical Sciences. Springer-Verlag, New York, 2002. [Tan79] H. Tanabe. Equations of Evolution. Pitman, London, 1979. [Tem83] R. Temam. Navier-Stokes Equations and Nonlinear Functional Analysis. CBMS-NSF Reg. Conf. Series in Appl. Math., SIAM, Philadelphia, 1983.

380

Bibliography [Tem88] R. Temam. Infinite-Dimensional Dynamical Systems in Mechanics and Physics, volume 68 of Applied Mathematical Sciences. SpringerVerlag, New York, 1988. [Tri95] H. Triebel. Interpolation Theory, Function Spaces, Differential Operators. J.A. Barth Verlag, Leipzig, 1995. [TY94] M. Taboado and Y. You. Invariant manifolds for retarded semilinear wave equations. J. Differ. Equations, 114(2):337–369, 1994. [Ver90] F. Verhulst. Nonlinear Ordinary Differential Equations and Dynamical Systems. Springer-Verlag, Berlin, 1990. [WF97] Z. Wang and X. Fan. Inertial manifolds for nonautonomous evolution equations. Adv. Math., Beijing, 26(2):185–186, 1997. [Yos80] K. Yosida. Functional Analysis. 6th ed., volume 123 of Grundlehren der mathematischen Wissenschaften. Springer-Verlag, Berlin, 1980. [Zei95] E. Zeidler. Applied Functional Analysis. Applications to Mathematical Physics. Vol. 1., volume 108 of Applied Mathematical Sciences. Springer-Verlag, Berlin, 1995. [ZM03] S. Zheng and A. Milani. Global attractors for singular perturbations of the Cahn-Hilliard equations. Preprint, 2003. [ZM05] S. Zheng and A. Milani. Exponential attractors and inertial manifolds for singular perturbations of the Cahn-Hilliard equations. To appear in Nonlinear Analysis - TMA, 2005.

Index

decay property, modified, 202 decomposition, 72, 73, 117 derivative, Fréchet, 344 derivative, weak, 354 diffeomorphism, 344 dimension, fractal, 81, 108, 166 dissipativity, xiii, 16 distance, Hausdorff, 42 distribution, regular, 353 dual, topological, 95 duality pairing, 329

attractor, xiii, 10, 17–20, 32, 33, 36, 38, 40, 41, 60, 63–65, 67, 69–71, 74, 79, 89, 91, 92, 107, 121, 123, 124, 128, 132, 133, 136, 152, 153, 159, 161, 162, 178, 193, 241, 269 attractor, butterfly, 36 attractor, compact, 33, 68–70, 73, 75, 143 attractor, continuity, 136 attractor, exponential, xiii, xv, xvi, 19, 20, 40, 64, 108, 135–137, 139, 140, 143, 146, 153, 162, 163, 175, 176, 183, 230, 241 attractor, finite dimensional, 93 attractor, global, xv, 65, 66, 68–70, 73, 78, 93, 107, 121, 178 attractor, lower semicontinuity, 132 attractor, maximal, 65 attractors, exponential, 206 attractors, upper semicontinuity, 132

equation, heat, xiii equation, autonomous, 40, 46 equation, beam, 93, 241 equation, Boussinesq, 33 equation, Burger, xiv, 135 equation, Chafee-Infante, xiv, 90, 92, 135, 177, 229 equation, dissipative, 73, 79, 89 equation, Duffing, xv, 3, 36, 38, 48, 87 equation, evolution, 4 equation, heat, 4, 71, 89 equation, hyperbolic, xvi, 73, 89, 92, 93, 111, 123, 132, 140, 147 equation, integral, 94 equation, Klein-Gordon, 90 equation, Korteweg-de Vries, xiv equation, Kuramoto-Sivashinski, xiv, 135 equation, logistic, 3, 37 equation, nondissipative, 93 equation, parabolic, xvi, 19, 67, 71, 89, 92, 93, 99, 132, 143, 147 equation, reaction-diffusion, 90, 92 equation, sequence, 30

basin of attraction, 64, 65 basis, total, 328 behavior, chaotic, 8, 10, 32 behavior, regular, 8 behavior, transient, 8, 9 bifurcation, 3 compactness, measure of, 74 completeness, asymptotic, 182 condition, spectral gap, 177, 179, 208, 210, 212, 220 cone property, 155 convergence, singular, 132

381

382 equation, sine-Gordon, 90, 92 equation, von Kármán, 93, 241, 272 equation, wave, xiii, xv, 72, 89 equations, Cahn-Hilliard, 92, 93, 135, 241 equations, Lorenz, xv, 3, 33, 46, 87, 93 equations, Maxwell, xiv, 10, 241, 280 equations, Navier-Stokes, xiv, 10, 92, 135, 241 Feigenbaum cascade, 3 flow, 43, 44, 56, 57, 59, 63, 114 form, inertial, 18, 179 formula, Duhamel, 94 function, absolutely continuous, 325 generator, infinitesimal, 346 graph transformation map, 186 graph transformation method, xvi, 185, 195, 197 group, 43 Hadamard, xvi homeomorphism, 344 horseshoe maps, 4 inequality, differential, 93, 106, 115, 120, 123, 144, 150 inequality, exponential, 86, 277 inequality, Friedrichs’, 365, 366 inequality, Gagliardo-Nirenberg, 356 inequality, Gronwall, 84, 102 inequality, Hölder, 350 inequality, interpolation, 105, 343 inequality, Minkowski, 351 inequality, Poincaré, 95, 98, 106, 359 inequality, Schwarz, 351 inequality, Young, 350 injection, compact, 336 lemma, Baire’s, 331 lemma, Lax-Milgram, 338 limit cycle, 327 manifold, finite dimensional, 18, 40

Index manifold, inertial, xiii, xv, xvi, 17–20, 40, 65, 135, 136, 139, 177, 178, 181, 241 manifold, stable, 34, 36, 39, 49, 50, 57, 59 manifold, unstable, 34, 37, 39, 49, 50, 57, 59 map, α-contracting, xv, 75, 77, 79, 80, 93, 108, 121, 123, 140, 250, 267, 269 map, circle-doubling, 24 map, differentiable, 344 map, horseshoe, 3 map, iterated, 1 map, logistic, 4 map, Poincaré, 23, 38 map, stroboscopic, 23, 38 map, tent, 28 mapping, norm-coercive, 344 motion, 5, 14, 132 operator, closed, 331 operator, coercive, 337 operator, compact, 332 operator, Laplace, 92, 95, 358 operator, positive, 337 operator, resolvent, 332 operator, self-adjoint, 335 operator, solution, 5, 106, 107 operator, strictly positive, 337 operator, symmetric, 335 operator, trace, 357 orbit, 6, 14, 50, 326 orbit, backward, 50 orbit, complete, 50, 51, 59 orbit, forward, 50 orbit, heteroclinic, 58 perturbation, 132 perturbation, singular, 90 phase, asymptotic, 182, 183 point spectrum, 332 point, asymptotically stable, 21, 56, 326 point, attractive, 21

Index point, equilibrium, 56, 326 point, fixed, 20 point, hyperbolic, 57 point, periodic, 21 point, stable, 20, 56, 326 point, stationary, 55, 326 point, unstable, 20, 56, 326 principle, slaving, 179 problem, Cauchy, 6, 12 process, dynamical, 1, 2, 5, 6, 9 projection, 137, 149 projection, orthogonal, 98, 138, 144, 153, 155 property, cone, 162 property, cone invariance, 177, 178, 189, 190, 203, 212 property, decay, 178, 193, 212 property, discrete squeezing, 135, 137, 139, 140, 143–145, 147, 149, 178 property, exponential tracking, 178, 182 property, modified decay, 203 property, regularity, 65 property, squeezing, 138–140, 153, 191 property, strong squeezing, 177, 178, 189, 193, 254 property, strong squeezing, modified, 203 pseudometric, 121, 123, 267, 269 pseudometric, precompact, 80, 121, 267 resolvent, compact, 333 section, Poincaré, 23 semicontinuity, 152 semicontinuity, lower, 133 semicontinuity, upper, 90, 132 semidistance, 42 semiflow, xiii–xv, 14, 15, 40–42, 44, 46, 48, 55, 57, 59–62, 64– 70, 72–74, 86, 89, 92, 93, 105, 107, 111, 114, 117,

383 121, 133, 135, 136, 139, 143, 144, 147, 149, 177, 178, 208, 241 semiflow, continuous, 43, 46, 135, 177, 181, 190, 191 semiflow, discrete, 43, 49, 139 semiflow, uniformly compact, 68 semigroup, 11, 43, 94, 177, 208, 345 semigroup, analytic, 90, 346 semigroup, continuous, 40, 94, 345 sequence, Bernoulli, 24 sequence, iterated, 20 sequence, Poincaré, 39 set, absorbing, 16, 68–71, 73, 77, 86, 93, 105, 106, 115, 118, 140, 143, 147, 175 set, attracting, 10, 16, 17, 61 set, fractal, 4, 16 set, inertial, xiii, 19 set, invariant, 17, 49, 50, 59, 62–64, 68, 71, 73, 153 set, limit, 49, 53–55, 60, 63 set, negatively invariant, 49, 59 set, positively invariant, 17, 49, 50, 59, 77, 86, 93, 105, 115, 122, 140, 143, 147, 153, 163, 164, 176 set, precompact, 61 set, relatively compact, 61 set, relatively dense, 371 set, resolvent, 332 set, totally bounded, 74 solution, generalized, 324 solution, mild, 94, 348 solution, periodic, 24 solution, strong, 348 solution, weak, 100, 101, 112, 113 space, interpolation, 343 space, phase, 8, 10, 14, 22, 39, 52, 90 system, autonomous, 11, 93, 325 system, chaotic, 10 system, complete orthogonal, 97, 359 system, discrete, 19 system, dissipative, 68, 69, 93, 140 system, dynamical, xiii, 40

384 system, finite dimensional, 40 system, uniformly compact, 67, 70, 72 system, well-posed, 140 test functions, 353 theorem, Bochner, 372 theorem, Carathéodory, 325 theorem, comparison, 84, 104 theorem, Hahn-Banach, 329 theorem, Hartman-Grobman, 326 theorem, Hille-Yosida, 346 theorem, Poincaré-Bendixon, 327 theorem, Stone, 347 well-posedness, 6, 47 well-posedness in the large, 6, 105, 114

Index

Nomenclature

AC(A, B) the space of the absolutely continuous functions from A to B, 325

CP

the cone in X with respect to the projection P, 138

Dα

derivative to the multiindex α, 353

α(Y)

the α-limit set of Y, 53

α(x)

the α-limit set of the point x, 55

diam(M) the diameter of M, 74

α(A)

the compactness measure of A, 74

dimF (K) the fractal dimension of K, 81

A0

the adjoint (or transpose) of the operator As, 333

As

a fractional power of A, 340

A∗

the adjoint of the operator A, 333

B(a, r) the ball with center a and radius r in X, 105 B(a, r) the closed ball with center a and radius r in X, 155 Bδ (a, r) the ball with center a and radius r with respect to the pseudometric δ , 80

d(a, b) the distance of the points a and b, 41 d(a, B) the distance of the point a to the set B, 41 ∂ (A, B) the semidistance from the A to the set B, 41 dist(A, B) the distance of the sets A and B, 41 D(Ω ) the space of the test functions on Ω , 353 D0 (Ω ) the space of the distributions on Ω , 353

C(A; B) the space of continuous functions from A into B, 5

E>a

the set {x ∈ E : x > a}, 41

C∞ 0 (Ω ) the space of the infinitely differentiable functions with compact support in Ω , 350

E≥a

the set {x ∈ E : x ≥ a}, 41

E
the set {x ∈ E : x < a}, 41

E≤a

the set {x ∈ E : x ≤ a}, 41

F0

the Fréchet derivative of F, 344

γ(x)

the complete orbit through x, 50

γ− (x)

the backward orbit starting at x, 50

C1 (A; B) the space of continuously differentiable functions from A into B, 5 CL

the cone in X with respect to the projections π1 , π2 and the parameter L, 189

385

386 γ+ (x)

Nomenclature the forward orbit starting at x, 50

Hm (Ω ) the Sobolev space of the functions on Ω having weak derivatives up to the order m in L2 (Ω ), 354 Hm (a, b; X ) the Sobolev space of the functions from ]a, b[ into X having weak derivatives up to the order m in L2 (a, b; X ), 362 Hs (Ω ) the interpolation [Hm (Ω ), L2 (Ω )]1−s/m m ≥ s, 355

space with

Hs0 (Ω ) the closure of C∞ 0 (Ω ) in Hs (Ω ), 357 ,→

the continuous injection, 336

bxc

the largest integer less than or equal to x, 24

[a, b[

the left closed, right open interval from a to b, 5

[a, b]

the closed interval from a to b, 19

p (Ω ) the space of the equivalence Lloc classes of locally Lebesgue pintegrable functions on Ω , 349

L p (a, b; X ) the space of the pintegrable functions on ]a, b[ with values in X , 361 L∞ (a, b; X ) the space of measurable, essentially bounded functions on ]a, b[ with values in X , 361 Ms (x) the stable manifold through x, 57 Mu (x) the unstable manifold through x, 57 N

the set of natural numbers (including 0), 2

ω(Y)

the ω-limit set of Y, 53

ω(x)

the ω-limit set of the point x, 55

R

the set of real numbers, xiv

ρ(A)

the resolvent set of A, 332

R(A, λ ) the resolvent of A, 332

]a, b[

the open interval from a to b, 5

σp (A) the point spectrum of A, 332

]a, b]

the left open, right closed interval from a to b, 21

T∗2

K(X , Y) the set of all linear, compact operators from X into Y, 332 L(X , Y) the set of all linear, bounded operators from X into Y, 330 L∞ (Ω ) the space of the equivalence classes of the measurable, essentially bounded functions on Ω , 349 L p (Ω ) the space of the equivalence classes of Lebesgue p-integrable functions on Ω , 349

the set {(t, τ) ∈ T × T : t ≥ τ}, 2

W(a, b; X , Y) for fixed m, the intermediate space of all functions in L2 (a, b; X ) whose m-th. derivative belongs to L2 (a, b; Y), 362 X0

the (topological) dual of X , 329

[X , Y]θ the interpolation space between X and Y to the parameter θ , 343 Z

the set of integer numbers, 2

An Introduction to Confucianism (Introduction to Religion)

An Introduction to Confucianism (Introduction to Religion)

An Introduction to Judaism (Introduction to Religion)

An Introduction to Judaism (Introduction to Religion)

An Introduction to Mormonism (Introduction to Religion)

An Introduction to Mormonism (Introduction to Religion)

An Introduction to Confucianism (Introduction to Religion)

An Introduction to Confucianism (Introduction to Religion)

An Introduction to Judaism (Introduction to Religion)

An Introduction to Judaism (Introduction to Religion)

An Introduction to Confucianism (Introduction to Religion)

An Introduction to Confucianism (Introduction to Religion)

An Introduction to Magnetohydrodynamics

An Introduction to Magnetohydrodynamics

An Introduction to Logic

An Introduction to Logic

An Introduction to Judaism

An Introduction to Judaism

An Introduction to Cryptography

An Introduction to Cryptography

An Introduction to Optimization

An Introduction to Optimization

An Introduction to Reasoning

An Introduction to Reasoning

An Introduction to Pharmacovigilance

An Introduction to Pharmacovigilance

An Introduction to Cybercultures

An Introduction to Cybercultures

An introduction to Hydrozoa

An introduction to Hydrozoa

An Introduction to Gerontology

An Introduction to Gerontology

An Introduction To Mechanics

An Introduction To Mechanics

An Introduction to Magnetohydrodynamics

An Introduction to Magnetohydrodynamics

An introduction to combinatorics

An introduction to combinatorics

An Introduction to Chemoinformatics

An Introduction to Chemoinformatics

An Introduction to

An Introduction to

An Introduction to Syntax

An Introduction to Syntax

An Introduction to Cryptography

An Introduction to Cryptography

An Introduction to Metaphysics

An Introduction to Metaphysics

An Introduction To Philosophy

An Introduction To Philosophy

An introduction to turbulence

An introduction to turbulence

An introduction to rheology

An introduction to rheology

An introduction to toxicogenomics

An introduction to toxicogenomics

An Introduction to Relativity

An Introduction to Relativity

An introduction to rights

An introduction to rights

Our partners will collect data and use cookies for ad personalization and measurement. Learn how we and our ad partner Google, collect and use data. Agree & close